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) 2013, 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
  38 #include <lustre/lustre_idl.h>
  39 #include <lu_object.h>
  40 #include <dt_object.h>
  41 #include <lustre_linkea.h>
  42 #include <lustre_fid.h>
  43 #include <lustre_lib.h>
  44 #include <lustre_net.h>
  45 #include <lustre/lustre_user.h>
  46 #include <md_object.h>
  47 #include <obd_class.h>
  48
  49 #include "lfsck_internal.h"
  50
  51 #define LFSCK_LAYOUT_MAGIC              0xB173AE14
  52
  53 static const char lfsck_layout_name[] = "lfsck_layout";
  54
  55 struct lfsck_layout_seq {
  56         struct list_head         lls_list;
  57         __u64                    lls_seq;
  58         __u64                    lls_lastid;
  59         __u64                    lls_lastid_known;
  60         struct dt_object        *lls_lastid_obj;
  61         unsigned int             lls_dirty:1;
  62 };
  63
  64 struct lfsck_layout_slave_data {
  65         /* list for lfsck_layout_seq */
  66         struct list_head         llsd_seq_list;
  67 };
  68
  69 struct lfsck_layout_object {
  70         struct dt_object        *llo_obj;
  71         struct lu_attr           llo_attr;
  72         atomic_t                 llo_ref;
  73         __u16                    llo_gen;
  74 };
  75
  76 struct lfsck_layout_req {
  77         struct list_head                 llr_list;
  78         struct lfsck_layout_object      *llr_parent;
  79         struct dt_object                *llr_child;
  80         __u32                            llr_ost_idx;
  81         __u32                            llr_lov_idx; /* offset in LOV EA */
  82 };
  83
  84 struct lfsck_layout_master_data {
  85         struct list_head        llmd_req_list;
  86         spinlock_t              llmd_lock;
  87         struct ptlrpc_thread    llmd_thread;
  88         atomic_t                llmd_rpcs_in_flight;
  89         int                     llmd_prefetched;
  90         int                     llmd_assistant_status;
  91         int                     llmd_post_result;
  92         unsigned int            llmd_to_post:1,
  93                                 llmd_to_double_scan:1,
  94                                 llmd_in_double_scan:1,
  95                                 llmd_exit:1;
  96 };
  97
  98 static inline void lfsck_layout_object_put(const struct lu_env *env,
  99                                            struct lfsck_layout_object *llo)
 100 {
 101         if (atomic_dec_and_test(&llo->llo_ref)) {
 102                 lfsck_object_put(env, llo->llo_obj);
 103                 OBD_FREE_PTR(llo);
 104         }
 105 }
 106
 107 static inline void lfsck_layout_req_fini(const struct lu_env *env,
 108                                          struct lfsck_layout_req *llr)
 109 {
 110         lu_object_put(env, &llr->llr_child->do_lu);
 111         lfsck_layout_object_put(env, llr->llr_parent);
 112         OBD_FREE_PTR(llr);
 113 }
 114
 115 static inline bool lfsck_layout_req_empty(struct lfsck_layout_master_data *llmd)
 116 {
 117         bool empty = false;
 118
 119         spin_lock(&llmd->llmd_lock);
 120         if (list_empty(&llmd->llmd_req_list))
 121                 empty = true;
 122         spin_unlock(&llmd->llmd_lock);
 123
 124         return empty;
 125 }
 126
 127 static void lfsck_layout_le_to_cpu(struct lfsck_layout *des,
 128                                    const struct lfsck_layout *src)
 129 {
 130         int i;
 131
 132         des->ll_magic = le32_to_cpu(src->ll_magic);
 133         des->ll_status = le32_to_cpu(src->ll_status);
 134         des->ll_flags = le32_to_cpu(src->ll_flags);
 135         des->ll_success_count = le32_to_cpu(src->ll_success_count);
 136         des->ll_run_time_phase1 = le32_to_cpu(src->ll_run_time_phase1);
 137         des->ll_run_time_phase2 = le32_to_cpu(src->ll_run_time_phase2);
 138         des->ll_time_last_complete = le64_to_cpu(src->ll_time_last_complete);
 139         des->ll_time_latest_start = le64_to_cpu(src->ll_time_latest_start);
 140         des->ll_time_last_checkpoint =
 141                                 le64_to_cpu(src->ll_time_last_checkpoint);
 142         des->ll_pos_latest_start = le64_to_cpu(src->ll_pos_latest_start);
 143         des->ll_pos_last_checkpoint = le64_to_cpu(src->ll_pos_last_checkpoint);
 144         des->ll_pos_first_inconsistent =
 145                         le64_to_cpu(src->ll_pos_first_inconsistent);
 146         des->ll_objs_checked_phase1 = le64_to_cpu(src->ll_objs_checked_phase1);
 147         des->ll_objs_failed_phase1 = le64_to_cpu(src->ll_objs_failed_phase1);
 148         des->ll_objs_checked_phase2 = le64_to_cpu(src->ll_objs_checked_phase2);
 149         des->ll_objs_failed_phase2 = le64_to_cpu(src->ll_objs_failed_phase2);
 150         for (i = 0; i < LLIT_MAX; i++)
 151                 des->ll_objs_repaired[i] =
 152                                 le64_to_cpu(src->ll_objs_repaired[i]);
 153         des->ll_objs_skipped = le64_to_cpu(src->ll_objs_skipped);
 154 }
 155
 156 static void lfsck_layout_cpu_to_le(struct lfsck_layout *des,
 157                                    const struct lfsck_layout *src)
 158 {
 159         int i;
 160
 161         des->ll_magic = cpu_to_le32(src->ll_magic);
 162         des->ll_status = cpu_to_le32(src->ll_status);
 163         des->ll_flags = cpu_to_le32(src->ll_flags);
 164         des->ll_success_count = cpu_to_le32(src->ll_success_count);
 165         des->ll_run_time_phase1 = cpu_to_le32(src->ll_run_time_phase1);
 166         des->ll_run_time_phase2 = cpu_to_le32(src->ll_run_time_phase2);
 167         des->ll_time_last_complete = cpu_to_le64(src->ll_time_last_complete);
 168         des->ll_time_latest_start = cpu_to_le64(src->ll_time_latest_start);
 169         des->ll_time_last_checkpoint =
 170                                 cpu_to_le64(src->ll_time_last_checkpoint);
 171         des->ll_pos_latest_start = cpu_to_le64(src->ll_pos_latest_start);
 172         des->ll_pos_last_checkpoint = cpu_to_le64(src->ll_pos_last_checkpoint);
 173         des->ll_pos_first_inconsistent =
 174                         cpu_to_le64(src->ll_pos_first_inconsistent);
 175         des->ll_objs_checked_phase1 = cpu_to_le64(src->ll_objs_checked_phase1);
 176         des->ll_objs_failed_phase1 = cpu_to_le64(src->ll_objs_failed_phase1);
 177         des->ll_objs_checked_phase2 = cpu_to_le64(src->ll_objs_checked_phase2);
 178         des->ll_objs_failed_phase2 = cpu_to_le64(src->ll_objs_failed_phase2);
 179         for (i = 0; i < LLIT_MAX; i++)
 180                 des->ll_objs_repaired[i] =
 181                                 cpu_to_le64(src->ll_objs_repaired[i]);
 182         des->ll_objs_skipped = cpu_to_le64(src->ll_objs_skipped);
 183 }
 184
 185 /**
 186  * \retval +ve: the lfsck_layout is broken, the caller should reset it.
 187  * \retval 0: succeed.
 188  * \retval -ve: failed cases.
 189  */
 190 static int lfsck_layout_load(const struct lu_env *env,
 191                              struct lfsck_component *com)
 192 {
 193         struct lfsck_layout             *lo     = com->lc_file_ram;
 194         const struct dt_body_operations *dbo    = com->lc_obj->do_body_ops;
 195         ssize_t                          size   = com->lc_file_size;
 196         loff_t                           pos    = 0;
 197         int                              rc;
 198
 199         rc = dbo->dbo_read(env, com->lc_obj,
 200                            lfsck_buf_get(env, com->lc_file_disk, size), &pos,
 201                            BYPASS_CAPA);
 202         if (rc == 0) {
 203                 return -ENOENT;
 204         } else if (rc < 0) {
 205                 CWARN("%s: failed to load lfsck_layout: rc = %d\n",
 206                       lfsck_lfsck2name(com->lc_lfsck), rc);
 207                 return rc;
 208         } else if (rc != size) {
 209                 CWARN("%s: crashed lfsck_layout, to be reset: rc = %d\n",
 210                       lfsck_lfsck2name(com->lc_lfsck), rc);
 211                 return 1;
 212         }
 213
 214         lfsck_layout_le_to_cpu(lo, com->lc_file_disk);
 215         if (lo->ll_magic != LFSCK_LAYOUT_MAGIC) {
 216                 CWARN("%s: invalid lfsck_layout magic %#x != %#x, "
 217                       "to be reset\n", lfsck_lfsck2name(com->lc_lfsck),
 218                       lo->ll_magic, LFSCK_LAYOUT_MAGIC);
 219                 return 1;
 220         }
 221
 222         return 0;
 223 }
 224
 225 static int lfsck_layout_store(const struct lu_env *env,
 226                               struct lfsck_component *com)
 227 {
 228         struct dt_object         *obj           = com->lc_obj;
 229         struct lfsck_instance    *lfsck         = com->lc_lfsck;
 230         struct lfsck_layout      *lo            = com->lc_file_disk;
 231         struct thandle           *handle;
 232         ssize_t                   size          = com->lc_file_size;
 233         loff_t                    pos           = 0;
 234         int                       rc;
 235         ENTRY;
 236
 237         lfsck_layout_cpu_to_le(lo, com->lc_file_ram);
 238         handle = dt_trans_create(env, lfsck->li_bottom);
 239         if (IS_ERR(handle)) {
 240                 rc = PTR_ERR(handle);
 241                 CERROR("%s: fail to create trans for storing lfsck_layout: "
 242                        "rc = %d\n", lfsck_lfsck2name(lfsck), rc);
 243                 RETURN(rc);
 244         }
 245
 246         rc = dt_declare_record_write(env, obj, size, pos, handle);
 247         if (rc != 0) {
 248                 CERROR("%s: fail to declare trans for storing lfsck_layout(1): "
 249                        "rc = %d\n", lfsck_lfsck2name(lfsck), rc);
 250                 GOTO(out, rc);
 251         }
 252
 253         rc = dt_trans_start_local(env, lfsck->li_bottom, handle);
 254         if (rc != 0) {
 255                 CERROR("%s: fail to start trans for storing lfsck_layout: "
 256                        "rc = %d\n", lfsck_lfsck2name(lfsck), rc);
 257                 GOTO(out, rc);
 258         }
 259
 260         rc = dt_record_write(env, obj, lfsck_buf_get(env, lo, size), &pos,
 261                              handle);
 262         if (rc != 0)
 263                 CERROR("%s: fail to store lfsck_layout(1): size = %d, "
 264                        "rc = %d\n", lfsck_lfsck2name(lfsck), (int)size, rc);
 265
 266         GOTO(out, rc);
 267
 268 out:
 269         dt_trans_stop(env, lfsck->li_bottom, handle);
 270
 271         return rc;
 272 }
 273
 274 static int lfsck_layout_init(const struct lu_env *env,
 275                              struct lfsck_component *com)
 276 {
 277         struct lfsck_layout *lo = com->lc_file_ram;
 278         int rc;
 279
 280         memset(lo, 0, com->lc_file_size);
 281         lo->ll_magic = LFSCK_LAYOUT_MAGIC;
 282         lo->ll_status = LS_INIT;
 283         down_write(&com->lc_sem);
 284         rc = lfsck_layout_store(env, com);
 285         up_write(&com->lc_sem);
 286
 287         return rc;
 288 }
 289
 290 static int fid_is_for_ostobj(const struct lu_env *env, struct dt_device *dt,
 291                              struct dt_object *obj, const struct lu_fid *fid)
 292 {
 293         struct seq_server_site  *ss     = lu_site2seq(dt->dd_lu_dev.ld_site);
 294         struct lu_seq_range      range  = { 0 };
 295         struct lustre_mdt_attrs *lma;
 296         int                      rc;
 297
 298         fld_range_set_any(&range);
 299         rc = fld_server_lookup(env, ss->ss_server_fld, fid_seq(fid), &range);
 300         if (rc == 0) {
 301                 if (fld_range_is_ost(&range))
 302                         return 1;
 303
 304                 return 0;
 305         }
 306
 307         lma = &lfsck_env_info(env)->lti_lma;
 308         rc = dt_xattr_get(env, obj, lfsck_buf_get(env, lma, sizeof(*lma)),
 309                           XATTR_NAME_LMA, BYPASS_CAPA);
 310         if (rc == sizeof(*lma)) {
 311                 lustre_lma_swab(lma);
 312
 313                 /* Generally, the low layer OSD create handler or OI scrub
 314                  * will set the LMAC_FID_ON_OST for all external visible
 315                  * OST-objects. But to make the otable-based iteration to
 316                  * be independent from OI scrub in spite of it got failure
 317                  * or not, we check the LMAC_FID_ON_OST here to guarantee
 318                  * that the LFSCK will not repair something by wrong. */
 319                 return lma->lma_compat & LMAC_FID_ON_OST ? 1 : 0;
 320         }
 321
 322         rc = dt_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_FID, BYPASS_CAPA);
 323
 324         return rc > 0;
 325 }
 326
 327 static struct lfsck_layout_seq *
 328 lfsck_layout_seq_lookup(struct lfsck_layout_slave_data *llsd, __u64 seq)
 329 {
 330         struct lfsck_layout_seq *lls;
 331
 332         list_for_each_entry(lls, &llsd->llsd_seq_list, lls_list) {
 333                 if (lls->lls_seq == seq)
 334                         return lls;
 335
 336                 if (lls->lls_seq > seq)
 337                         return NULL;
 338         }
 339
 340         return NULL;
 341 }
 342
 343 static void
 344 lfsck_layout_seq_insert(struct lfsck_layout_slave_data *llsd,
 345                         struct lfsck_layout_seq *lls)
 346 {
 347         struct lfsck_layout_seq *tmp;
 348         struct list_head        *pos = &llsd->llsd_seq_list;
 349
 350         list_for_each_entry(tmp, &llsd->llsd_seq_list, lls_list) {
 351                 if (lls->lls_seq < tmp->lls_seq) {
 352                         pos = &tmp->lls_list;
 353                         break;
 354                 }
 355         }
 356         list_add_tail(&lls->lls_list, pos);
 357 }
 358
 359 static int
 360 lfsck_layout_lastid_create(const struct lu_env *env,
 361                            struct lfsck_instance *lfsck,
 362                            struct dt_object *obj)
 363 {
 364         struct lfsck_thread_info *info   = lfsck_env_info(env);
 365         struct lu_attr           *la     = &info->lti_la;
 366         struct dt_object_format  *dof    = &info->lti_dof;
 367         struct lfsck_bookmark    *bk     = &lfsck->li_bookmark_ram;
 368         struct dt_device         *dt     = lfsck->li_bottom;
 369         struct thandle           *th;
 370         __u64                     lastid = 0;
 371         loff_t                    pos    = 0;
 372         int                       rc;
 373         ENTRY;
 374
 375         CDEBUG(D_LFSCK, "To create LAST_ID for <seq> "LPX64"\n",
 376                fid_seq(lfsck_dto2fid(obj)));
 377
 378         if (bk->lb_param & LPF_DRYRUN)
 379                 return 0;
 380
 381         memset(la, 0, sizeof(*la));
 382         la->la_mode = S_IFREG |  S_IRUGO | S_IWUSR;
 383         la->la_valid = LA_MODE | LA_UID | LA_GID;
 384         dof->dof_type = dt_mode_to_dft(S_IFREG);
 385
 386         th = dt_trans_create(env, dt);
 387         if (IS_ERR(th))
 388                 RETURN(rc = PTR_ERR(th));
 389
 390         rc = dt_declare_create(env, obj, la, NULL, dof, th);
 391         if (rc != 0)
 392                 GOTO(stop, rc);
 393
 394         rc = dt_declare_record_write(env, obj, sizeof(lastid), pos, th);
 395         if (rc != 0)
 396                 GOTO(stop, rc);
 397
 398         rc = dt_trans_start_local(env, dt, th);
 399         if (rc != 0)
 400                 GOTO(stop, rc);
 401
 402         dt_write_lock(env, obj, 0);
 403         if (likely(!dt_object_exists(obj))) {
 404                 rc = dt_create(env, obj, la, NULL, dof, th);
 405                 if (rc == 0)
 406                         rc = dt_record_write(env, obj,
 407                                 lfsck_buf_get(env, &lastid, sizeof(lastid)),
 408                                 &pos, th);
 409         }
 410         dt_write_unlock(env, obj);
 411
 412         GOTO(stop, rc);
 413
 414 stop:
 415         dt_trans_stop(env, dt, th);
 416
 417         return rc;
 418 }
 419
 420 static int
 421 lfsck_layout_lastid_reload(const struct lu_env *env,
 422                            struct lfsck_component *com,
 423                            struct lfsck_layout_seq *lls)
 424 {
 425         __u64   lastid;
 426         loff_t  pos     = 0;
 427         int     rc;
 428
 429         dt_read_lock(env, lls->lls_lastid_obj, 0);
 430         rc = dt_record_read(env, lls->lls_lastid_obj,
 431                             lfsck_buf_get(env, &lastid, sizeof(lastid)), &pos);
 432         dt_read_unlock(env, lls->lls_lastid_obj);
 433         if (unlikely(rc != 0))
 434                 return rc;
 435
 436         lastid = le64_to_cpu(lastid);
 437         if (lastid < lls->lls_lastid_known) {
 438                 struct lfsck_instance   *lfsck  = com->lc_lfsck;
 439                 struct lfsck_layout     *lo     = com->lc_file_ram;
 440
 441                 lls->lls_lastid = lls->lls_lastid_known;
 442                 lls->lls_dirty = 1;
 443                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
 444                         LASSERT(lfsck->li_out_notify != NULL);
 445
 446                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
 447                                              LE_LASTID_REBUILDING);
 448                         lo->ll_flags |= LF_CRASHED_LASTID;
 449                 }
 450         } else if (lastid >= lls->lls_lastid) {
 451                 lls->lls_lastid = lastid;
 452                 lls->lls_dirty = 0;
 453         }
 454
 455         return 0;
 456 }
 457
 458 static int
 459 lfsck_layout_lastid_store(const struct lu_env *env,
 460                           struct lfsck_component *com)
 461 {
 462         struct lfsck_instance           *lfsck  = com->lc_lfsck;
 463         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
 464         struct dt_device                *dt     = lfsck->li_bottom;
 465         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
 466         struct lfsck_layout_seq         *lls;
 467         struct thandle                  *th;
 468         __u64                            lastid;
 469         int                              rc     = 0;
 470         int                              rc1    = 0;
 471
 472         list_for_each_entry(lls, &llsd->llsd_seq_list, lls_list) {
 473                 loff_t pos = 0;
 474
 475                 /* XXX: Add the code back if we really found related
 476                  *      inconsistent cases in the future. */
 477 #if 0
 478                 if (!lls->lls_dirty) {
 479                         /* In OFD, before the pre-creation, the LAST_ID
 480                          * file will be updated firstly, which may hide
 481                          * some potential crashed cases. For example:
 482                          *
 483                          * The old obj1's ID is higher than old LAST_ID
 484                          * but lower than the new LAST_ID, but the LFSCK
 485                          * have not touch the obj1 until the OFD updated
 486                          * the LAST_ID. So the LFSCK does not regard it
 487                          * as crashed case. But when OFD does not create
 488                          * successfully, it will set the LAST_ID as the
 489                          * real created objects' ID, then LFSCK needs to
 490                          * found related inconsistency. */
 491                         rc = lfsck_layout_lastid_reload(env, com, lls);
 492                         if (likely(!lls->lls_dirty))
 493                                 continue;
 494                 }
 495 #endif
 496
 497                 CDEBUG(D_LFSCK, "To sync the LAST_ID for <seq> "LPX64
 498                        " as <oid> "LPU64"\n", lls->lls_seq, lls->lls_lastid);
 499
 500                 if (bk->lb_param & LPF_DRYRUN) {
 501                         lls->lls_dirty = 0;
 502                         continue;
 503                 }
 504
 505                 th = dt_trans_create(env, dt);
 506                 if (IS_ERR(th)) {
 507                         rc1 = PTR_ERR(th);
 508                         CERROR("%s: (1) failed to store "LPX64": rc = %d\n",
 509                                lfsck_lfsck2name(com->lc_lfsck),
 510                                lls->lls_seq, rc1);
 511                         continue;
 512                 }
 513
 514                 rc = dt_declare_record_write(env, lls->lls_lastid_obj,
 515                                              sizeof(lastid), pos, th);
 516                 if (rc != 0)
 517                         goto stop;
 518
 519                 rc = dt_trans_start_local(env, dt, th);
 520                 if (rc != 0)
 521                         goto stop;
 522
 523                 lastid = cpu_to_le64(lls->lls_lastid);
 524                 dt_write_lock(env, lls->lls_lastid_obj, 0);
 525                 rc = dt_record_write(env, lls->lls_lastid_obj,
 526                                      lfsck_buf_get(env, &lastid,
 527                                      sizeof(lastid)), &pos, th);
 528                 dt_write_unlock(env, lls->lls_lastid_obj);
 529                 if (rc == 0)
 530                         lls->lls_dirty = 0;
 531
 532 stop:
 533                 dt_trans_stop(env, dt, th);
 534                 if (rc != 0) {
 535                         rc1 = rc;
 536                         CERROR("%s: (2) failed to store "LPX64": rc = %d\n",
 537                                lfsck_lfsck2name(com->lc_lfsck),
 538                                lls->lls_seq, rc1);
 539                 }
 540         }
 541
 542         return rc1;
 543 }
 544
 545 static int
 546 lfsck_layout_lastid_load(const struct lu_env *env,
 547                          struct lfsck_component *com,
 548                          struct lfsck_layout_seq *lls)
 549 {
 550         struct lfsck_instance   *lfsck  = com->lc_lfsck;
 551         struct lfsck_layout     *lo     = com->lc_file_ram;
 552         struct lu_fid           *fid    = &lfsck_env_info(env)->lti_fid;
 553         struct dt_object        *obj;
 554         loff_t                   pos    = 0;
 555         int                      rc;
 556         ENTRY;
 557
 558         lu_last_id_fid(fid, lls->lls_seq, lfsck_dev_idx(lfsck->li_bottom));
 559         obj = dt_locate(env, lfsck->li_bottom, fid);
 560         if (IS_ERR(obj))
 561                 RETURN(PTR_ERR(obj));
 562
 563         /* LAST_ID crashed, to be rebuilt */
 564         if (!dt_object_exists(obj)) {
 565                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
 566                         LASSERT(lfsck->li_out_notify != NULL);
 567
 568                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
 569                                              LE_LASTID_REBUILDING);
 570                         lo->ll_flags |= LF_CRASHED_LASTID;
 571
 572                         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DELAY4) &&
 573                             cfs_fail_val > 0) {
 574                                 struct l_wait_info lwi = LWI_TIMEOUT(
 575                                                 cfs_time_seconds(cfs_fail_val),
 576                                                 NULL, NULL);
 577
 578                                 up_write(&com->lc_sem);
 579                                 l_wait_event(lfsck->li_thread.t_ctl_waitq,
 580                                              !thread_is_running(&lfsck->li_thread),
 581                                              &lwi);
 582                                 down_write(&com->lc_sem);
 583                         }
 584                 }
 585
 586                 rc = lfsck_layout_lastid_create(env, lfsck, obj);
 587         } else {
 588                 dt_read_lock(env, obj, 0);
 589                 rc = dt_read(env, obj,
 590                         lfsck_buf_get(env, &lls->lls_lastid, sizeof(__u64)),
 591                         &pos);
 592                 dt_read_unlock(env, obj);
 593                 if (rc != 0 && rc != sizeof(__u64))
 594                         GOTO(out, rc = (rc > 0 ? -EFAULT : rc));
 595
 596                 if (rc == 0 && !(lo->ll_flags & LF_CRASHED_LASTID)) {
 597                         LASSERT(lfsck->li_out_notify != NULL);
 598
 599                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
 600                                              LE_LASTID_REBUILDING);
 601                         lo->ll_flags |= LF_CRASHED_LASTID;
 602                 }
 603
 604                 lls->lls_lastid = le64_to_cpu(lls->lls_lastid);
 605                 rc = 0;
 606         }
 607
 608         GOTO(out, rc);
 609
 610 out:
 611         if (rc != 0)
 612                 lfsck_object_put(env, obj);
 613         else
 614                 lls->lls_lastid_obj = obj;
 615
 616         return rc;
 617 }
 618
 619 static int lfsck_layout_master_query_others(const struct lu_env *env,
 620                                             struct lfsck_component *com)
 621 {
 622         /* XXX: to be implemented. */
 623
 624         return 0;
 625 }
 626
 627 static inline bool
 628 lfsck_layout_master_to_orphan(struct lfsck_layout_master_data *llmd)
 629 {
 630         /* XXX: to be implemented. */
 631
 632         return 1;
 633 }
 634
 635 static int lfsck_layout_master_notify_others(const struct lu_env *env,
 636                                              struct lfsck_component *com,
 637                                              struct lfsck_request *lr)
 638 {
 639         /* XXX: to be implemented. */
 640
 641         return 0;
 642 }
 643
 644 static int lfsck_layout_double_scan_result(const struct lu_env *env,
 645                                            struct lfsck_component *com,
 646                                            int rc)
 647 {
 648         struct lfsck_instance   *lfsck = com->lc_lfsck;
 649         struct lfsck_layout     *lo    = com->lc_file_ram;
 650         struct lfsck_bookmark   *bk    = &lfsck->li_bookmark_ram;
 651
 652         down_write(&com->lc_sem);
 653
 654         lo->ll_run_time_phase2 += cfs_duration_sec(cfs_time_current() +
 655                                 HALF_SEC - lfsck->li_time_last_checkpoint);
 656         lo->ll_time_last_checkpoint = cfs_time_current_sec();
 657         lo->ll_objs_checked_phase2 += com->lc_new_checked;
 658
 659         if (rc > 0) {
 660                 com->lc_journal = 0;
 661                 if (lo->ll_flags & LF_INCOMPLETE)
 662                         lo->ll_status = LS_PARTIAL;
 663                 else
 664                         lo->ll_status = LS_COMPLETED;
 665                 if (!(bk->lb_param & LPF_DRYRUN))
 666                         lo->ll_flags &= ~(LF_SCANNED_ONCE | LF_INCONSISTENT);
 667                 lo->ll_time_last_complete = lo->ll_time_last_checkpoint;
 668                 lo->ll_success_count++;
 669         } else if (rc == 0) {
 670                 lo->ll_status = lfsck->li_status;
 671                 if (lo->ll_status == 0)
 672                         lo->ll_status = LS_STOPPED;
 673         } else {
 674                 lo->ll_status = LS_FAILED;
 675         }
 676
 677         if (lo->ll_status != LS_PAUSED) {
 678                 spin_lock(&lfsck->li_lock);
 679                 list_del_init(&com->lc_link);
 680                 list_add_tail(&com->lc_link, &lfsck->li_list_idle);
 681                 spin_unlock(&lfsck->li_lock);
 682         }
 683
 684         rc = lfsck_layout_store(env, com);
 685
 686         up_write(&com->lc_sem);
 687
 688         return rc;
 689 }
 690
 691 static int lfsck_layout_assistant(void *args)
 692 {
 693         struct lfsck_thread_args        *lta     = args;
 694         struct lu_env                   *env     = &lta->lta_env;
 695         struct lfsck_component          *com     = lta->lta_com;
 696         struct lfsck_instance           *lfsck   = lta->lta_lfsck;
 697         struct lfsck_bookmark           *bk      = &lfsck->li_bookmark_ram;
 698         struct lfsck_position           *pos     = &com->lc_pos_start;
 699         struct lfsck_thread_info        *info    = lfsck_env_info(env);
 700         struct lfsck_request            *lr      = &info->lti_lr;
 701         struct lfsck_layout_master_data *llmd    = com->lc_data;
 702         struct ptlrpc_thread            *mthread = &lfsck->li_thread;
 703         struct ptlrpc_thread            *athread = &llmd->llmd_thread;
 704         struct lfsck_layout_req         *llr;
 705         struct l_wait_info               lwi     = { 0 };
 706         int                              rc      = 0;
 707         int                              rc1     = 0;
 708         ENTRY;
 709
 710         memset(lr, 0, sizeof(*lr));
 711         lr->lr_event = LE_START;
 712         lr->lr_index = lfsck_dev_idx(lfsck->li_bottom);
 713         lr->lr_valid = LSV_SPEED_LIMIT | LSV_ERROR_HANDLE | LSV_DRYRUN |
 714                        LSV_ASYNC_WINDOWS;
 715         lr->lr_speed = bk->lb_speed_limit;
 716         lr->lr_version = bk->lb_version;
 717         lr->lr_active = LT_LAYOUT;
 718         lr->lr_param = bk->lb_param;
 719         lr->lr_async_windows = bk->lb_async_windows;
 720         if (pos->lp_oit_cookie <= 1)
 721                 lr->lr_param |= LPF_RESET;
 722
 723         rc = lfsck_layout_master_notify_others(env, com, lr);
 724         if (rc != 0) {
 725                 CERROR("%s: fail to notify others for layout start: rc = %d\n",
 726                        lfsck_lfsck2name(lfsck), rc);
 727                 GOTO(fini, rc);
 728         }
 729
 730         spin_lock(&llmd->llmd_lock);
 731         thread_set_flags(athread, SVC_RUNNING);
 732         spin_unlock(&llmd->llmd_lock);
 733         wake_up_all(&mthread->t_ctl_waitq);
 734
 735         while (1) {
 736                 while (!list_empty(&llmd->llmd_req_list)) {
 737                         bool wakeup = false;
 738
 739                         l_wait_event(athread->t_ctl_waitq,
 740                                      bk->lb_async_windows == 0 ||
 741                                      atomic_read(&llmd->llmd_rpcs_in_flight) <
 742                                                 bk->lb_async_windows ||
 743                                      llmd->llmd_exit,
 744                                      &lwi);
 745
 746                         if (unlikely(llmd->llmd_exit))
 747                                 GOTO(cleanup1, rc = llmd->llmd_post_result);
 748
 749                         /* XXX: To be extended in other patch.
 750                          *
 751                          * Compare the OST side attribute with local attribute,
 752                          * and fix it if found inconsistency. */
 753
 754                         spin_lock(&llmd->llmd_lock);
 755                         llr = list_entry(llmd->llmd_req_list.next,
 756                                          struct lfsck_layout_req,
 757                                          llr_list);
 758                         list_del_init(&llr->llr_list);
 759                         if (bk->lb_async_windows != 0 &&
 760                             llmd->llmd_prefetched >= bk->lb_async_windows)
 761                                 wakeup = true;
 762
 763                         llmd->llmd_prefetched--;
 764                         spin_unlock(&llmd->llmd_lock);
 765                         if (wakeup)
 766                                 wake_up_all(&mthread->t_ctl_waitq);
 767
 768                         lfsck_layout_req_fini(env, llr);
 769                 }
 770
 771                 /* Wakeup the master engine if it is waiting in checkpoint. */
 772                 if (atomic_read(&llmd->llmd_rpcs_in_flight) == 0)
 773                         wake_up_all(&mthread->t_ctl_waitq);
 774
 775                 l_wait_event(athread->t_ctl_waitq,
 776                              !lfsck_layout_req_empty(llmd) ||
 777                              llmd->llmd_exit ||
 778                              llmd->llmd_to_post ||
 779                              llmd->llmd_to_double_scan,
 780                              &lwi);
 781
 782                 if (unlikely(llmd->llmd_exit))
 783                         GOTO(cleanup1, rc = llmd->llmd_post_result);
 784
 785                 if (!list_empty(&llmd->llmd_req_list))
 786                         continue;
 787
 788                 if (llmd->llmd_to_post) {
 789                         llmd->llmd_to_post = 0;
 790                         LASSERT(llmd->llmd_post_result > 0);
 791
 792                         memset(lr, 0, sizeof(*lr));
 793                         lr->lr_index = lfsck_dev_idx(lfsck->li_bottom);
 794                         lr->lr_active = LT_LAYOUT;
 795                         lr->lr_event = LE_PHASE1_DONE;
 796                         lr->lr_status = llmd->llmd_post_result;
 797                         rc = lfsck_layout_master_notify_others(env, com, lr);
 798                         if (rc != 0)
 799                                 CERROR("%s: failed to notify others "
 800                                        "for layout post: rc = %d\n",
 801                                        lfsck_lfsck2name(lfsck), rc);
 802
 803                         /* Wakeup the master engine to go ahead. */
 804                         wake_up_all(&mthread->t_ctl_waitq);
 805                 }
 806
 807                 if (llmd->llmd_to_double_scan) {
 808                         llmd->llmd_to_double_scan = 0;
 809                         atomic_inc(&lfsck->li_double_scan_count);
 810                         llmd->llmd_in_double_scan = 1;
 811                         wake_up_all(&mthread->t_ctl_waitq);
 812
 813                         while (llmd->llmd_in_double_scan) {
 814                                 rc = lfsck_layout_master_query_others(env, com);
 815                                 if (lfsck_layout_master_to_orphan(llmd))
 816                                         goto orphan;
 817
 818                                 if (rc < 0)
 819                                         GOTO(cleanup2, rc);
 820
 821                                 /* Pull LFSCK status on related targets once
 822                                  * per 30 seconds if we are not notified. */
 823                                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(30),
 824                                                            cfs_time_seconds(1),
 825                                                            NULL, NULL);
 826                                 rc = l_wait_event(athread->t_ctl_waitq,
 827                                         lfsck_layout_master_to_orphan(llmd) ||
 828                                         llmd->llmd_exit ||
 829                                         !thread_is_running(mthread),
 830                                         &lwi);
 831
 832                                 if (unlikely(llmd->llmd_exit ||
 833                                              !thread_is_running(mthread)))
 834                                         GOTO(cleanup2, rc = 0);
 835
 836                                 if (rc == -ETIMEDOUT)
 837                                         continue;
 838
 839                                 if (rc < 0)
 840                                         GOTO(cleanup2, rc);
 841
 842 orphan:
 843                                 /* XXX: real double scan for ost orphans. */
 844
 845                                 GOTO(cleanup2, rc = 1);
 846                         }
 847                 }
 848         }
 849
 850 cleanup1:
 851         /* Cleanup the unfinished requests. */
 852         spin_lock(&llmd->llmd_lock);
 853         while (!list_empty(&llmd->llmd_req_list)) {
 854                 llr = list_entry(llmd->llmd_req_list.next,
 855                                  struct lfsck_layout_req,
 856                                  llr_list);
 857                 list_del_init(&llr->llr_list);
 858                 llmd->llmd_prefetched--;
 859                 spin_unlock(&llmd->llmd_lock);
 860                 lfsck_layout_req_fini(env, llr);
 861                 spin_lock(&llmd->llmd_lock);
 862         }
 863         spin_unlock(&llmd->llmd_lock);
 864
 865         LASSERTF(llmd->llmd_prefetched == 0, "unmatched prefeteched objs %d\n",
 866                  llmd->llmd_prefetched);
 867
 868         l_wait_event(athread->t_ctl_waitq,
 869                      atomic_read(&llmd->llmd_rpcs_in_flight) == 0,
 870                      &lwi);
 871
 872 cleanup2:
 873         memset(lr, 0, sizeof(*lr));
 874         lr->lr_index = lfsck_dev_idx(lfsck->li_bottom);
 875         lr->lr_active = LT_LAYOUT;
 876         if (rc > 0) {
 877                 lr->lr_event = LE_PHASE2_DONE;
 878                 lr->lr_status = rc;
 879         } else if (rc == 0) {
 880                 lr->lr_event = LE_STOP;
 881                 if (lfsck->li_status == LS_PAUSED ||
 882                     lfsck->li_status == LS_CO_PAUSED)
 883                         lr->lr_status = LS_CO_PAUSED;
 884                 else if (lfsck->li_status == LS_STOPPED ||
 885                          lfsck->li_status == LS_CO_STOPPED)
 886                         lr->lr_status = LS_CO_STOPPED;
 887                 else
 888                         LBUG();
 889         } else {
 890                 lr->lr_event = LE_STOP;
 891                 lr->lr_status = LS_CO_FAILED;
 892         }
 893
 894         rc1 = lfsck_layout_master_notify_others(env, com, lr);
 895         if (rc1 != 0) {
 896                 CERROR("%s: failed to notify others for layout quit: rc = %d\n",
 897                        lfsck_lfsck2name(lfsck), rc1);
 898                 rc = rc1;
 899         }
 900
 901         /* Under force exit case, some requests may be just freed without
 902          * verification, those objects should be re-handled when next run.
 903          * So not update the on-disk tracing file under such case. */
 904         if (!llmd->llmd_exit)
 905                 rc1 = lfsck_layout_double_scan_result(env, com, rc);
 906
 907 fini:
 908         if (llmd->llmd_in_double_scan)
 909                 atomic_dec(&lfsck->li_double_scan_count);
 910
 911         spin_lock(&llmd->llmd_lock);
 912         llmd->llmd_assistant_status = (rc1 != 0 ? rc1 : rc);
 913         thread_set_flags(athread, SVC_STOPPED);
 914         wake_up_all(&mthread->t_ctl_waitq);
 915         spin_unlock(&llmd->llmd_lock);
 916         lfsck_thread_args_fini(lta);
 917
 918         return rc;
 919 }
 920
 921 /* layout APIs */
 922
 923 static int lfsck_layout_reset(const struct lu_env *env,
 924                               struct lfsck_component *com, bool init)
 925 {
 926         struct lfsck_layout     *lo    = com->lc_file_ram;
 927         int                      rc;
 928
 929         down_write(&com->lc_sem);
 930         if (init) {
 931                 memset(lo, 0, com->lc_file_size);
 932         } else {
 933                 __u32 count = lo->ll_success_count;
 934                 __u64 last_time = lo->ll_time_last_complete;
 935
 936                 memset(lo, 0, com->lc_file_size);
 937                 lo->ll_success_count = count;
 938                 lo->ll_time_last_complete = last_time;
 939         }
 940
 941         lo->ll_magic = LFSCK_LAYOUT_MAGIC;
 942         lo->ll_status = LS_INIT;
 943
 944         rc = lfsck_layout_store(env, com);
 945         up_write(&com->lc_sem);
 946
 947         return rc;
 948 }
 949
 950 static void lfsck_layout_fail(const struct lu_env *env,
 951                               struct lfsck_component *com, bool new_checked)
 952 {
 953         struct lfsck_layout *lo = com->lc_file_ram;
 954
 955         down_write(&com->lc_sem);
 956         if (new_checked)
 957                 com->lc_new_checked++;
 958         lo->ll_objs_failed_phase1++;
 959         if (lo->ll_pos_first_inconsistent == 0) {
 960                 struct lfsck_instance *lfsck = com->lc_lfsck;
 961
 962                 lo->ll_pos_first_inconsistent =
 963                         lfsck->li_obj_oit->do_index_ops->dio_it.store(env,
 964                                                         lfsck->li_di_oit);
 965         }
 966         up_write(&com->lc_sem);
 967 }
 968
 969 static int lfsck_layout_master_checkpoint(const struct lu_env *env,
 970                                           struct lfsck_component *com, bool init)
 971 {
 972         struct lfsck_instance           *lfsck   = com->lc_lfsck;
 973         struct lfsck_layout             *lo      = com->lc_file_ram;
 974         struct lfsck_layout_master_data *llmd    = com->lc_data;
 975         struct ptlrpc_thread            *mthread = &lfsck->li_thread;
 976         struct ptlrpc_thread            *athread = &llmd->llmd_thread;
 977         struct l_wait_info               lwi     = { 0 };
 978         int                              rc;
 979
 980         if (com->lc_new_checked == 0 && !init)
 981                 return 0;
 982
 983         l_wait_event(mthread->t_ctl_waitq,
 984                      (list_empty(&llmd->llmd_req_list) &&
 985                       atomic_read(&llmd->llmd_rpcs_in_flight) == 0) ||
 986                      !thread_is_running(mthread) ||
 987                      thread_is_stopped(athread),
 988                      &lwi);
 989
 990         if (!thread_is_running(mthread) || thread_is_stopped(athread))
 991                 return 0;
 992
 993         down_write(&com->lc_sem);
 994         if (init) {
 995                 lo->ll_pos_latest_start = lfsck->li_pos_current.lp_oit_cookie;
 996         } else {
 997                 lo->ll_pos_last_checkpoint =
 998                                         lfsck->li_pos_current.lp_oit_cookie;
 999                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
1000                                 HALF_SEC - lfsck->li_time_last_checkpoint);
1001                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
1002                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
1003                 com->lc_new_checked = 0;
1004         }
1005
1006         rc = lfsck_layout_store(env, com);
1007         up_write(&com->lc_sem);
1008
1009         return rc;
1010 }
1011
1012 static int lfsck_layout_slave_checkpoint(const struct lu_env *env,
1013                                          struct lfsck_component *com, bool init)
1014 {
1015         struct lfsck_instance   *lfsck = com->lc_lfsck;
1016         struct lfsck_layout     *lo    = com->lc_file_ram;
1017         int                      rc;
1018
1019         if (com->lc_new_checked == 0 && !init)
1020                 return 0;
1021
1022         down_write(&com->lc_sem);
1023
1024         if (init) {
1025                 lo->ll_pos_latest_start = lfsck->li_pos_current.lp_oit_cookie;
1026         } else {
1027                 lo->ll_pos_last_checkpoint =
1028                                         lfsck->li_pos_current.lp_oit_cookie;
1029                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
1030                                 HALF_SEC - lfsck->li_time_last_checkpoint);
1031                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
1032                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
1033                 com->lc_new_checked = 0;
1034         }
1035
1036         rc = lfsck_layout_store(env, com);
1037
1038         up_write(&com->lc_sem);
1039
1040         return rc;
1041 }
1042
1043 static int lfsck_layout_slave_prep(const struct lu_env *env,
1044                                    struct lfsck_component *com)
1045 {
1046         struct lfsck_instance   *lfsck  = com->lc_lfsck;
1047         struct lfsck_layout     *lo     = com->lc_file_ram;
1048         struct lfsck_position   *pos    = &com->lc_pos_start;
1049
1050         /* XXX: For a new scanning, generate OST-objects
1051          *      bitmap for orphan detection. */
1052
1053         fid_zero(&pos->lp_dir_parent);
1054         pos->lp_dir_cookie = 0;
1055         if (lo->ll_status == LS_COMPLETED ||
1056             lo->ll_status == LS_PARTIAL) {
1057                 int rc;
1058
1059                 rc = lfsck_layout_reset(env, com, false);
1060                 if (rc != 0)
1061                         return rc;
1062         }
1063
1064         down_write(&com->lc_sem);
1065
1066         lo->ll_time_latest_start = cfs_time_current_sec();
1067
1068         spin_lock(&lfsck->li_lock);
1069         if (lo->ll_flags & LF_SCANNED_ONCE) {
1070                 if (!lfsck->li_drop_dryrun ||
1071                     lo->ll_pos_first_inconsistent == 0) {
1072                         lo->ll_status = LS_SCANNING_PHASE2;
1073                         list_del_init(&com->lc_link);
1074                         list_add_tail(&com->lc_link,
1075                                       &lfsck->li_list_double_scan);
1076                         pos->lp_oit_cookie = 0;
1077                 } else {
1078                         int i;
1079
1080                         lo->ll_status = LS_SCANNING_PHASE1;
1081                         lo->ll_run_time_phase1 = 0;
1082                         lo->ll_run_time_phase2 = 0;
1083                         lo->ll_objs_checked_phase1 = 0;
1084                         lo->ll_objs_checked_phase2 = 0;
1085                         lo->ll_objs_failed_phase1 = 0;
1086                         lo->ll_objs_failed_phase2 = 0;
1087                         for (i = 0; i < LLIT_MAX; i++)
1088                                 lo->ll_objs_repaired[i] = 0;
1089
1090                         pos->lp_oit_cookie = lo->ll_pos_first_inconsistent;
1091                 }
1092         } else {
1093                 lo->ll_status = LS_SCANNING_PHASE1;
1094                 if (!lfsck->li_drop_dryrun ||
1095                     lo->ll_pos_first_inconsistent == 0)
1096                         pos->lp_oit_cookie = lo->ll_pos_last_checkpoint + 1;
1097                 else
1098                         pos->lp_oit_cookie = lo->ll_pos_first_inconsistent;
1099         }
1100         spin_unlock(&lfsck->li_lock);
1101
1102         up_write(&com->lc_sem);
1103
1104         return 0;
1105 }
1106
1107 static int lfsck_layout_master_prep(const struct lu_env *env,
1108                                     struct lfsck_component *com)
1109 {
1110         struct lfsck_instance           *lfsck   = com->lc_lfsck;
1111         struct lfsck_layout_master_data *llmd    = com->lc_data;
1112         struct ptlrpc_thread            *mthread = &lfsck->li_thread;
1113         struct ptlrpc_thread            *athread = &llmd->llmd_thread;
1114         struct lfsck_thread_args        *lta;
1115         long                             rc;
1116         ENTRY;
1117
1118         rc = lfsck_layout_slave_prep(env, com);
1119         if (rc != 0)
1120                 RETURN(rc);
1121
1122         llmd->llmd_assistant_status = 0;
1123         llmd->llmd_post_result = 0;
1124         llmd->llmd_to_post = 0;
1125         llmd->llmd_to_double_scan = 0;
1126         llmd->llmd_in_double_scan = 0;
1127         llmd->llmd_exit = 0;
1128         thread_set_flags(athread, 0);
1129
1130         lta = lfsck_thread_args_init(lfsck, com);
1131         if (IS_ERR(lta))
1132                 RETURN(PTR_ERR(lta));
1133
1134         rc = PTR_ERR(kthread_run(lfsck_layout_assistant, lta, "lfsck_layout"));
1135         if (IS_ERR_VALUE(rc)) {
1136                 CERROR("%s: Cannot start LFSCK layout assistant thread: "
1137                        "rc = %ld\n", lfsck_lfsck2name(lfsck), rc);
1138                 lfsck_thread_args_fini(lta);
1139         } else {
1140                 struct l_wait_info lwi = { 0 };
1141
1142                 l_wait_event(mthread->t_ctl_waitq,
1143                              thread_is_running(athread) ||
1144                              thread_is_stopped(athread),
1145                              &lwi);
1146                 if (unlikely(!thread_is_running(athread)))
1147                         rc = llmd->llmd_assistant_status;
1148                 else
1149                         rc = 0;
1150         }
1151
1152         RETURN(rc);
1153 }
1154
1155 static int lfsck_layout_master_exec_oit(const struct lu_env *env,
1156                                         struct lfsck_component *com,
1157                                         struct dt_object *obj)
1158 {
1159         /* XXX: To be implemented in other patches.
1160          *
1161          * For the given object, read its layout EA locally. For each stripe,
1162          * pre-fetch the OST-object's attribute and generate an structure
1163          * lfsck_layout_req on the list ::llmd_req_list.
1164          *
1165          * For each request on the ::llmd_req_list, the lfsck_layout_assistant
1166          * thread will compare the OST side attribute with local attribute,
1167          * if inconsistent, then repair it.
1168          *
1169          * All above processing is async mode with pipeline. */
1170
1171         return 0;
1172 }
1173
1174 static int lfsck_layout_slave_exec_oit(const struct lu_env *env,
1175                                        struct lfsck_component *com,
1176                                        struct dt_object *obj)
1177 {
1178         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1179         struct lfsck_layout             *lo     = com->lc_file_ram;
1180         const struct lu_fid             *fid    = lfsck_dto2fid(obj);
1181         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
1182         struct lfsck_layout_seq         *lls;
1183         __u64                            seq;
1184         __u64                            oid;
1185         int                              rc;
1186         ENTRY;
1187
1188         /* XXX: Update OST-objects bitmap for orphan detection. */
1189
1190         LASSERT(llsd != NULL);
1191
1192         down_write(&com->lc_sem);
1193         if (fid_is_idif(fid))
1194                 seq = 0;
1195         else if (!fid_is_norm(fid) ||
1196                  !fid_is_for_ostobj(env, lfsck->li_next, obj, fid))
1197                 GOTO(unlock, rc = 0);
1198         else
1199                 seq = fid_seq(fid);
1200         com->lc_new_checked++;
1201
1202         lls = lfsck_layout_seq_lookup(llsd, seq);
1203         if (lls == NULL) {
1204                 OBD_ALLOC_PTR(lls);
1205                 if (unlikely(lls == NULL))
1206                         GOTO(unlock, rc = -ENOMEM);
1207
1208                 INIT_LIST_HEAD(&lls->lls_list);
1209                 lls->lls_seq = seq;
1210                 rc = lfsck_layout_lastid_load(env, com, lls);
1211                 if (rc != 0) {
1212                         lo->ll_objs_failed_phase1++;
1213                         OBD_FREE_PTR(lls);
1214                         GOTO(unlock, rc);
1215                 }
1216
1217                 lfsck_layout_seq_insert(llsd, lls);
1218         }
1219
1220         if (unlikely(fid_is_last_id(fid)))
1221                 GOTO(unlock, rc = 0);
1222
1223         oid = fid_oid(fid);
1224         if (oid > lls->lls_lastid_known)
1225                 lls->lls_lastid_known = oid;
1226
1227         if (oid > lls->lls_lastid) {
1228                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
1229                         /* OFD may create new objects during LFSCK scanning. */
1230                         rc = lfsck_layout_lastid_reload(env, com, lls);
1231                         if (unlikely(rc != 0))
1232                                 CWARN("%s: failed to reload LAST_ID for "LPX64
1233                                       ": rc = %d\n",
1234                                       lfsck_lfsck2name(com->lc_lfsck),
1235                                       lls->lls_seq, rc);
1236                         if (oid <= lls->lls_lastid)
1237                                 GOTO(unlock, rc = 0);
1238
1239                         LASSERT(lfsck->li_out_notify != NULL);
1240
1241                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1242                                              LE_LASTID_REBUILDING);
1243                         lo->ll_flags |= LF_CRASHED_LASTID;
1244                 }
1245
1246                 lls->lls_lastid = oid;
1247                 lls->lls_dirty = 1;
1248         }
1249
1250         GOTO(unlock, rc = 0);
1251
1252 unlock:
1253         up_write(&com->lc_sem);
1254
1255         return rc;
1256 }
1257
1258 static int lfsck_layout_exec_dir(const struct lu_env *env,
1259                                  struct lfsck_component *com,
1260                                  struct dt_object *obj,
1261                                  struct lu_dirent *ent)
1262 {
1263         return 0;
1264 }
1265
1266 static int lfsck_layout_master_post(const struct lu_env *env,
1267                                     struct lfsck_component *com,
1268                                     int result, bool init)
1269 {
1270         struct lfsck_instance           *lfsck   = com->lc_lfsck;
1271         struct lfsck_layout             *lo      = com->lc_file_ram;
1272         struct lfsck_layout_master_data *llmd    = com->lc_data;
1273         struct ptlrpc_thread            *mthread = &lfsck->li_thread;
1274         struct ptlrpc_thread            *athread = &llmd->llmd_thread;
1275         struct l_wait_info               lwi     = { 0 };
1276         int                              rc;
1277         ENTRY;
1278
1279
1280         llmd->llmd_post_result = result;
1281         llmd->llmd_to_post = 1;
1282         if (llmd->llmd_post_result <= 0)
1283                 llmd->llmd_exit = 1;
1284
1285         wake_up_all(&athread->t_ctl_waitq);
1286         l_wait_event(mthread->t_ctl_waitq,
1287                      (result > 0 && list_empty(&llmd->llmd_req_list) &&
1288                       atomic_read(&llmd->llmd_rpcs_in_flight) == 0) ||
1289                      thread_is_stopped(athread),
1290                      &lwi);
1291
1292         if (llmd->llmd_assistant_status < 0)
1293                 result = llmd->llmd_assistant_status;
1294
1295         down_write(&com->lc_sem);
1296         spin_lock(&lfsck->li_lock);
1297         /* When LFSCK failed, there may be some prefetched objects those are
1298          * not been processed yet, we do not know the exactly position, then
1299          * just restart from last check-point next time. */
1300         if (!init && !llmd->llmd_exit)
1301                 lo->ll_pos_last_checkpoint =
1302                                         lfsck->li_pos_current.lp_oit_cookie;
1303
1304         if (result > 0) {
1305                 lo->ll_status = LS_SCANNING_PHASE2;
1306                 lo->ll_flags |= LF_SCANNED_ONCE;
1307                 lo->ll_flags &= ~LF_UPGRADE;
1308                 list_del_init(&com->lc_link);
1309                 list_add_tail(&com->lc_link, &lfsck->li_list_double_scan);
1310         } else if (result == 0) {
1311                 lo->ll_status = lfsck->li_status;
1312                 if (lo->ll_status == 0)
1313                         lo->ll_status = LS_STOPPED;
1314                 if (lo->ll_status != LS_PAUSED) {
1315                         list_del_init(&com->lc_link);
1316                         list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1317                 }
1318         } else {
1319                 lo->ll_status = LS_FAILED;
1320                 list_del_init(&com->lc_link);
1321                 list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1322         }
1323         spin_unlock(&lfsck->li_lock);
1324
1325         if (!init) {
1326                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
1327                                 HALF_SEC - lfsck->li_time_last_checkpoint);
1328                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
1329                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
1330                 com->lc_new_checked = 0;
1331         }
1332
1333         rc = lfsck_layout_store(env, com);
1334         up_write(&com->lc_sem);
1335
1336         RETURN(rc);
1337 }
1338
1339 static int lfsck_layout_slave_post(const struct lu_env *env,
1340                                    struct lfsck_component *com,
1341                                    int result, bool init)
1342 {
1343         struct lfsck_instance   *lfsck = com->lc_lfsck;
1344         struct lfsck_layout     *lo    = com->lc_file_ram;
1345         int                      rc;
1346         bool                     done  = false;
1347
1348         rc = lfsck_layout_lastid_store(env, com);
1349         if (rc != 0)
1350                 result = rc;
1351
1352         LASSERT(lfsck->li_out_notify != NULL);
1353
1354         down_write(&com->lc_sem);
1355
1356         spin_lock(&lfsck->li_lock);
1357         if (!init)
1358                 lo->ll_pos_last_checkpoint =
1359                                         lfsck->li_pos_current.lp_oit_cookie;
1360         if (result > 0) {
1361                 lo->ll_status = LS_SCANNING_PHASE2;
1362                 lo->ll_flags |= LF_SCANNED_ONCE;
1363                 if (lo->ll_flags & LF_CRASHED_LASTID) {
1364                         done = true;
1365                         lo->ll_flags &= ~LF_CRASHED_LASTID;
1366                 }
1367                 lo->ll_flags &= ~LF_UPGRADE;
1368                 list_del_init(&com->lc_link);
1369                 list_add_tail(&com->lc_link, &lfsck->li_list_double_scan);
1370         } else if (result == 0) {
1371                 lo->ll_status = lfsck->li_status;
1372                 if (lo->ll_status == 0)
1373                         lo->ll_status = LS_STOPPED;
1374                 if (lo->ll_status != LS_PAUSED) {
1375                         list_del_init(&com->lc_link);
1376                         list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1377                 }
1378         } else {
1379                 lo->ll_status = LS_FAILED;
1380                 list_del_init(&com->lc_link);
1381                 list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1382         }
1383         spin_unlock(&lfsck->li_lock);
1384
1385         if (done)
1386                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1387                                      LE_LASTID_REBUILT);
1388
1389         if (!init) {
1390                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
1391                                 HALF_SEC - lfsck->li_time_last_checkpoint);
1392                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
1393                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
1394                 com->lc_new_checked = 0;
1395         }
1396
1397         rc = lfsck_layout_store(env, com);
1398
1399         up_write(&com->lc_sem);
1400
1401         return rc;
1402 }
1403
1404 static int lfsck_layout_dump(const struct lu_env *env,
1405                              struct lfsck_component *com, char *buf, int len)
1406 {
1407         struct lfsck_instance   *lfsck = com->lc_lfsck;
1408         struct lfsck_bookmark   *bk    = &lfsck->li_bookmark_ram;
1409         struct lfsck_layout     *lo    = com->lc_file_ram;
1410         int                      save  = len;
1411         int                      ret   = -ENOSPC;
1412         int                      rc;
1413
1414         down_read(&com->lc_sem);
1415         rc = snprintf(buf, len,
1416                       "name: lfsck_layout\n"
1417                       "magic: %#x\n"
1418                       "version: %d\n"
1419                       "status: %s\n",
1420                       lo->ll_magic,
1421                       bk->lb_version,
1422                       lfsck_status2names(lo->ll_status));
1423         if (rc <= 0)
1424                 goto out;
1425
1426         buf += rc;
1427         len -= rc;
1428         rc = lfsck_bits_dump(&buf, &len, lo->ll_flags, lfsck_flags_names,
1429                              "flags");
1430         if (rc < 0)
1431                 goto out;
1432
1433         rc = lfsck_bits_dump(&buf, &len, bk->lb_param, lfsck_param_names,
1434                              "param");
1435         if (rc < 0)
1436                 goto out;
1437
1438         rc = lfsck_time_dump(&buf, &len, lo->ll_time_last_complete,
1439                              "time_since_last_completed");
1440         if (rc < 0)
1441                 goto out;
1442
1443         rc = lfsck_time_dump(&buf, &len, lo->ll_time_latest_start,
1444                              "time_since_latest_start");
1445         if (rc < 0)
1446                 goto out;
1447
1448         rc = lfsck_time_dump(&buf, &len, lo->ll_time_last_checkpoint,
1449                              "time_since_last_checkpoint");
1450         if (rc < 0)
1451                 goto out;
1452
1453         rc = snprintf(buf, len,
1454                       "latest_start_position: "LPU64"\n"
1455                       "last_checkpoint_position: "LPU64"\n"
1456                       "first_failure_position: "LPU64"\n",
1457                       lo->ll_pos_latest_start,
1458                       lo->ll_pos_last_checkpoint,
1459                       lo->ll_pos_first_inconsistent);
1460         if (rc <= 0)
1461                 goto out;
1462
1463         buf += rc;
1464         len -= rc;
1465
1466         rc = snprintf(buf, len,
1467                       "success_count: %u\n"
1468                       "repaired_dangling: "LPU64"\n"
1469                       "repaired_unmatched_pair: "LPU64"\n"
1470                       "repaired_multiple_referenced: "LPU64"\n"
1471                       "repaired_orphan: "LPU64"\n"
1472                       "repaired_inconsistent_owner: "LPU64"\n"
1473                       "repaired_others: "LPU64"\n"
1474                       "skipped: "LPU64"\n"
1475                       "failed_phase1: "LPU64"\n"
1476                       "failed_phase2: "LPU64"\n",
1477                       lo->ll_success_count,
1478                       lo->ll_objs_repaired[LLIT_DANGLING - 1],
1479                       lo->ll_objs_repaired[LLIT_UNMATCHED_PAIR - 1],
1480                       lo->ll_objs_repaired[LLIT_MULTIPLE_REFERENCED - 1],
1481                       lo->ll_objs_repaired[LLIT_ORPHAN - 1],
1482                       lo->ll_objs_repaired[LLIT_INCONSISTENT_OWNER - 1],
1483                       lo->ll_objs_repaired[LLIT_OTHERS - 1],
1484                       lo->ll_objs_skipped,
1485                       lo->ll_objs_failed_phase1,
1486                       lo->ll_objs_failed_phase2);
1487         if (rc <= 0)
1488                 goto out;
1489
1490         buf += rc;
1491         len -= rc;
1492
1493         if (lo->ll_status == LS_SCANNING_PHASE1) {
1494                 __u64 pos;
1495                 const struct dt_it_ops *iops;
1496                 cfs_duration_t duration = cfs_time_current() -
1497                                           lfsck->li_time_last_checkpoint;
1498                 __u64 checked = lo->ll_objs_checked_phase1 + com->lc_new_checked;
1499                 __u64 speed = checked;
1500                 __u64 new_checked = com->lc_new_checked * HZ;
1501                 __u32 rtime = lo->ll_run_time_phase1 +
1502                               cfs_duration_sec(duration + HALF_SEC);
1503
1504                 if (duration != 0)
1505                         do_div(new_checked, duration);
1506                 if (rtime != 0)
1507                         do_div(speed, rtime);
1508                 rc = snprintf(buf, len,
1509                               "checked_phase1: "LPU64"\n"
1510                               "checked_phase2: "LPU64"\n"
1511                               "run_time_phase1: %u seconds\n"
1512                               "run_time_phase2: %u seconds\n"
1513                               "average_speed_phase1: "LPU64" items/sec\n"
1514                               "average_speed_phase2: N/A\n"
1515                               "real-time_speed_phase1: "LPU64" items/sec\n"
1516                               "real-time_speed_phase2: N/A\n",
1517                               checked,
1518                               lo->ll_objs_checked_phase2,
1519                               rtime,
1520                               lo->ll_run_time_phase2,
1521                               speed,
1522                               new_checked);
1523                 if (rc <= 0)
1524                         goto out;
1525
1526                 buf += rc;
1527                 len -= rc;
1528
1529                 LASSERT(lfsck->li_di_oit != NULL);
1530
1531                 iops = &lfsck->li_obj_oit->do_index_ops->dio_it;
1532
1533                 /* The low layer otable-based iteration position may NOT
1534                  * exactly match the layout-based directory traversal
1535                  * cookie. Generally, it is not a serious issue. But the
1536                  * caller should NOT make assumption on that. */
1537                 pos = iops->store(env, lfsck->li_di_oit);
1538                 if (!lfsck->li_current_oit_processed)
1539                         pos--;
1540                 rc = snprintf(buf, len, "current_position: "LPU64"\n", pos);
1541                 if (rc <= 0)
1542                         goto out;
1543
1544                 buf += rc;
1545                 len -= rc;
1546         } else {
1547                 /* XXX: LS_SCANNING_PHASE2 will be handled in the future. */
1548                 __u64 speed1 = lo->ll_objs_checked_phase1;
1549                 __u64 speed2 = lo->ll_objs_checked_phase2;
1550
1551                 if (lo->ll_run_time_phase1 != 0)
1552                         do_div(speed1, lo->ll_run_time_phase1);
1553                 if (lo->ll_run_time_phase2 != 0)
1554                         do_div(speed2, lo->ll_run_time_phase2);
1555                 rc = snprintf(buf, len,
1556                               "checked_phase1: "LPU64"\n"
1557                               "checked_phase2: "LPU64"\n"
1558                               "run_time_phase1: %u seconds\n"
1559                               "run_time_phase2: %u seconds\n"
1560                               "average_speed_phase1: "LPU64" items/sec\n"
1561                               "average_speed_phase2: "LPU64" objs/sec\n"
1562                               "real-time_speed_phase1: N/A\n"
1563                               "real-time_speed_phase2: N/A\n"
1564                               "current_position: N/A\n",
1565                               lo->ll_objs_checked_phase1,
1566                               lo->ll_objs_checked_phase2,
1567                               lo->ll_run_time_phase1,
1568                               lo->ll_run_time_phase2,
1569                               speed1,
1570                               speed2);
1571                 if (rc <= 0)
1572                         goto out;
1573
1574                 buf += rc;
1575                 len -= rc;
1576         }
1577         ret = save - len;
1578
1579 out:
1580         up_read(&com->lc_sem);
1581
1582         return ret;
1583 }
1584
1585 static int lfsck_layout_master_double_scan(const struct lu_env *env,
1586                                            struct lfsck_component *com)
1587 {
1588         struct lfsck_layout_master_data *llmd    = com->lc_data;
1589         struct ptlrpc_thread            *mthread = &com->lc_lfsck->li_thread;
1590         struct ptlrpc_thread            *athread = &llmd->llmd_thread;
1591         struct lfsck_layout             *lo      = com->lc_file_ram;
1592         struct l_wait_info               lwi     = { 0 };
1593
1594         if (unlikely(lo->ll_status != LS_SCANNING_PHASE2))
1595                 return 0;
1596
1597         llmd->llmd_to_double_scan = 1;
1598         wake_up_all(&athread->t_ctl_waitq);
1599         l_wait_event(mthread->t_ctl_waitq,
1600                      llmd->llmd_in_double_scan ||
1601                      thread_is_stopped(athread),
1602                      &lwi);
1603         if (llmd->llmd_assistant_status < 0)
1604                 return llmd->llmd_assistant_status;
1605
1606         return 0;
1607 }
1608
1609 static int lfsck_layout_slave_double_scan(const struct lu_env *env,
1610                                           struct lfsck_component *com)
1611 {
1612         struct lfsck_instance   *lfsck = com->lc_lfsck;
1613         struct lfsck_layout     *lo    = com->lc_file_ram;
1614         int                      rc    = 1;
1615
1616         if (unlikely(lo->ll_status != LS_SCANNING_PHASE2))
1617                 return 0;
1618
1619         atomic_inc(&lfsck->li_double_scan_count);
1620
1621         com->lc_new_checked = 0;
1622         com->lc_new_scanned = 0;
1623         com->lc_time_last_checkpoint = cfs_time_current();
1624         com->lc_time_next_checkpoint = com->lc_time_last_checkpoint +
1625                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
1626
1627         rc = lfsck_layout_double_scan_result(env, com, rc);
1628
1629         if (atomic_dec_and_test(&lfsck->li_double_scan_count))
1630                 wake_up_all(&lfsck->li_thread.t_ctl_waitq);
1631
1632         return rc;
1633 }
1634
1635 static void lfsck_layout_master_data_release(const struct lu_env *env,
1636                                              struct lfsck_component *com)
1637 {
1638         struct lfsck_layout_master_data *llmd = com->lc_data;
1639
1640         LASSERT(llmd != NULL);
1641         LASSERT(thread_is_init(&llmd->llmd_thread) ||
1642                 thread_is_stopped(&llmd->llmd_thread));
1643         LASSERT(list_empty(&llmd->llmd_req_list));
1644         LASSERT(atomic_read(&llmd->llmd_rpcs_in_flight) == 0);
1645
1646         com->lc_data = NULL;
1647         OBD_FREE_PTR(llmd);
1648 }
1649
1650 static void lfsck_layout_slave_data_release(const struct lu_env *env,
1651                                             struct lfsck_component *com)
1652 {
1653         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
1654         struct lfsck_layout_seq         *lls;
1655         struct lfsck_layout_seq         *next;
1656
1657         LASSERT(llsd != NULL);
1658
1659         com->lc_data = NULL;
1660
1661         list_for_each_entry_safe(lls, next, &llsd->llsd_seq_list,
1662                                      lls_list) {
1663                 list_del_init(&lls->lls_list);
1664                 lfsck_object_put(env, lls->lls_lastid_obj);
1665                 OBD_FREE_PTR(lls);
1666         }
1667
1668         OBD_FREE_PTR(llsd);
1669 }
1670
1671 static void lfsck_layout_master_quit(const struct lu_env *env,
1672                                      struct lfsck_component *com)
1673 {
1674         struct lfsck_layout_master_data *llmd    = com->lc_data;
1675         struct ptlrpc_thread            *mthread = &com->lc_lfsck->li_thread;
1676         struct ptlrpc_thread            *athread = &llmd->llmd_thread;
1677         struct l_wait_info               lwi     = { 0 };
1678
1679         llmd->llmd_exit = 1;
1680         wake_up_all(&athread->t_ctl_waitq);
1681         l_wait_event(mthread->t_ctl_waitq,
1682                      thread_is_init(athread) ||
1683                      thread_is_stopped(athread),
1684                      &lwi);
1685 }
1686
1687 static int lfsck_layout_master_in_notify(const struct lu_env *env,
1688                                          struct lfsck_component *com,
1689                                          struct lfsck_request *lr)
1690 {
1691         /* XXX: to record the event from layout slave on the OST. */
1692         return 0;
1693 }
1694
1695 static int lfsck_layout_slave_in_notify(const struct lu_env *env,
1696                                         struct lfsck_component *com,
1697                                         struct lfsck_request *lr)
1698 {
1699         /* XXX: to record the event from layout master on the MDT. */
1700         return 0;
1701 }
1702
1703 static int lfsck_layout_query(const struct lu_env *env,
1704                               struct lfsck_component *com)
1705 {
1706         struct lfsck_layout *lo = com->lc_file_ram;
1707
1708         return lo->ll_status;
1709 }
1710
1711 static struct lfsck_operations lfsck_layout_master_ops = {
1712         .lfsck_reset            = lfsck_layout_reset,
1713         .lfsck_fail             = lfsck_layout_fail,
1714         .lfsck_checkpoint       = lfsck_layout_master_checkpoint,
1715         .lfsck_prep             = lfsck_layout_master_prep,
1716         .lfsck_exec_oit         = lfsck_layout_master_exec_oit,
1717         .lfsck_exec_dir         = lfsck_layout_exec_dir,
1718         .lfsck_post             = lfsck_layout_master_post,
1719         .lfsck_dump             = lfsck_layout_dump,
1720         .lfsck_double_scan      = lfsck_layout_master_double_scan,
1721         .lfsck_data_release     = lfsck_layout_master_data_release,
1722         .lfsck_quit             = lfsck_layout_master_quit,
1723         .lfsck_in_notify        = lfsck_layout_master_in_notify,
1724         .lfsck_query            = lfsck_layout_query,
1725 };
1726
1727 static struct lfsck_operations lfsck_layout_slave_ops = {
1728         .lfsck_reset            = lfsck_layout_reset,
1729         .lfsck_fail             = lfsck_layout_fail,
1730         .lfsck_checkpoint       = lfsck_layout_slave_checkpoint,
1731         .lfsck_prep             = lfsck_layout_slave_prep,
1732         .lfsck_exec_oit         = lfsck_layout_slave_exec_oit,
1733         .lfsck_exec_dir         = lfsck_layout_exec_dir,
1734         .lfsck_post             = lfsck_layout_slave_post,
1735         .lfsck_dump             = lfsck_layout_dump,
1736         .lfsck_double_scan      = lfsck_layout_slave_double_scan,
1737         .lfsck_data_release     = lfsck_layout_slave_data_release,
1738         .lfsck_in_notify        = lfsck_layout_slave_in_notify,
1739         .lfsck_query            = lfsck_layout_query,
1740 };
1741
1742 int lfsck_layout_setup(const struct lu_env *env, struct lfsck_instance *lfsck)
1743 {
1744         struct lfsck_component  *com;
1745         struct lfsck_layout     *lo;
1746         struct dt_object        *root = NULL;
1747         struct dt_object        *obj;
1748         int                      rc;
1749         ENTRY;
1750
1751         OBD_ALLOC_PTR(com);
1752         if (com == NULL)
1753                 RETURN(-ENOMEM);
1754
1755         INIT_LIST_HEAD(&com->lc_link);
1756         INIT_LIST_HEAD(&com->lc_link_dir);
1757         init_rwsem(&com->lc_sem);
1758         atomic_set(&com->lc_ref, 1);
1759         com->lc_lfsck = lfsck;
1760         com->lc_type = LT_LAYOUT;
1761         if (lfsck->li_master) {
1762                 struct lfsck_layout_master_data *llmd;
1763
1764                 com->lc_ops = &lfsck_layout_master_ops;
1765                 OBD_ALLOC_PTR(llmd);
1766                 if (llmd == NULL)
1767                         GOTO(out, rc = -ENOMEM);
1768
1769                 INIT_LIST_HEAD(&llmd->llmd_req_list);
1770                 spin_lock_init(&llmd->llmd_lock);
1771                 init_waitqueue_head(&llmd->llmd_thread.t_ctl_waitq);
1772                 atomic_set(&llmd->llmd_rpcs_in_flight, 0);
1773                 com->lc_data = llmd;
1774         } else {
1775                 struct lfsck_layout_slave_data *llsd;
1776
1777                 com->lc_ops = &lfsck_layout_slave_ops;
1778                 OBD_ALLOC_PTR(llsd);
1779                 if (llsd == NULL)
1780                         GOTO(out, rc = -ENOMEM);
1781
1782                 INIT_LIST_HEAD(&llsd->llsd_seq_list);
1783                 com->lc_data = llsd;
1784         }
1785         com->lc_file_size = sizeof(*lo);
1786         OBD_ALLOC(com->lc_file_ram, com->lc_file_size);
1787         if (com->lc_file_ram == NULL)
1788                 GOTO(out, rc = -ENOMEM);
1789
1790         OBD_ALLOC(com->lc_file_disk, com->lc_file_size);
1791         if (com->lc_file_disk == NULL)
1792                 GOTO(out, rc = -ENOMEM);
1793
1794         root = dt_locate(env, lfsck->li_bottom, &lfsck->li_local_root_fid);
1795         if (IS_ERR(root))
1796                 GOTO(out, rc = PTR_ERR(root));
1797
1798         if (unlikely(!dt_try_as_dir(env, root)))
1799                 GOTO(out, rc = -ENOTDIR);
1800
1801         obj = local_file_find_or_create(env, lfsck->li_los, root,
1802                                         lfsck_layout_name,
1803                                         S_IFREG | S_IRUGO | S_IWUSR);
1804         if (IS_ERR(obj))
1805                 GOTO(out, rc = PTR_ERR(obj));
1806
1807         com->lc_obj = obj;
1808         rc = lfsck_layout_load(env, com);
1809         if (rc > 0)
1810                 rc = lfsck_layout_reset(env, com, true);
1811         else if (rc == -ENOENT)
1812                 rc = lfsck_layout_init(env, com);
1813
1814         if (rc != 0)
1815                 GOTO(out, rc);
1816
1817         lo = com->lc_file_ram;
1818         switch (lo->ll_status) {
1819         case LS_INIT:
1820         case LS_COMPLETED:
1821         case LS_FAILED:
1822         case LS_STOPPED:
1823         case LS_PARTIAL:
1824                 spin_lock(&lfsck->li_lock);
1825                 list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1826                 spin_unlock(&lfsck->li_lock);
1827                 break;
1828         default:
1829                 CERROR("%s: unknown lfsck_layout status: rc = %u\n",
1830                        lfsck_lfsck2name(lfsck), lo->ll_status);
1831                 /* fall through */
1832         case LS_SCANNING_PHASE1:
1833         case LS_SCANNING_PHASE2:
1834                 /* No need to store the status to disk right now.
1835                  * If the system crashed before the status stored,
1836                  * it will be loaded back when next time. */
1837                 lo->ll_status = LS_CRASHED;
1838                 lo->ll_flags |= LF_INCOMPLETE;
1839                 /* fall through */
1840         case LS_PAUSED:
1841         case LS_CRASHED:
1842         case LS_CO_FAILED:
1843         case LS_CO_STOPPED:
1844         case LS_CO_PAUSED:
1845                 spin_lock(&lfsck->li_lock);
1846                 list_add_tail(&com->lc_link, &lfsck->li_list_scan);
1847                 spin_unlock(&lfsck->li_lock);
1848                 break;
1849         }
1850
1851         if (lo->ll_flags & LF_CRASHED_LASTID) {
1852                 LASSERT(lfsck->li_out_notify != NULL);
1853
1854                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1855                                      LE_LASTID_REBUILDING);
1856         }
1857
1858         GOTO(out, rc = 0);
1859
1860 out:
1861         if (root != NULL && !IS_ERR(root))
1862                 lu_object_put(env, &root->do_lu);
1863
1864         if (rc != 0)
1865                 lfsck_component_cleanup(env, com);
1866
1867         return rc;
1868 }