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                 if (lfsck->li_paused)
 671                         lo->ll_status = LS_PAUSED;
 672                 else
 673                         lo->ll_status = LS_STOPPED;
 674         } else {
 675                 lo->ll_status = LS_FAILED;
 676         }
 677
 678         if (lo->ll_status != LS_PAUSED) {
 679                 spin_lock(&lfsck->li_lock);
 680                 list_del_init(&com->lc_link);
 681                 list_add_tail(&com->lc_link, &lfsck->li_list_idle);
 682                 spin_unlock(&lfsck->li_lock);
 683         }
 684
 685         rc = lfsck_layout_store(env, com);
 686
 687         up_write(&com->lc_sem);
 688
 689         return rc;
 690 }
 691
 692 static int lfsck_layout_assistant(void *args)
 693 {
 694         struct lfsck_thread_args        *lta     = args;
 695         struct lu_env                   *env     = &lta->lta_env;
 696         struct lfsck_component          *com     = lta->lta_com;
 697         struct lfsck_instance           *lfsck   = lta->lta_lfsck;
 698         struct lfsck_bookmark           *bk      = &lfsck->li_bookmark_ram;
 699         struct lfsck_position           *pos     = &com->lc_pos_start;
 700         struct lfsck_thread_info        *info    = lfsck_env_info(env);
 701         struct lfsck_request            *lr      = &info->lti_lr;
 702         struct lfsck_layout_master_data *llmd    = com->lc_data;
 703         struct ptlrpc_thread            *mthread = &lfsck->li_thread;
 704         struct ptlrpc_thread            *athread = &llmd->llmd_thread;
 705         struct lfsck_layout_req         *llr;
 706         struct l_wait_info               lwi     = { 0 };
 707         int                              rc      = 0;
 708         int                              rc1     = 0;
 709         ENTRY;
 710
 711         memset(lr, 0, sizeof(*lr));
 712         lr->lr_event = LE_START;
 713         lr->lr_index = lfsck_dev_idx(lfsck->li_bottom);
 714         lr->lr_valid = LSV_SPEED_LIMIT | LSV_ERROR_HANDLE | LSV_DRYRUN |
 715                        LSV_ASYNC_WINDOWS;
 716         lr->lr_speed = bk->lb_speed_limit;
 717         lr->lr_version = bk->lb_version;
 718         lr->lr_active = LT_LAYOUT;
 719         lr->lr_param = bk->lb_param;
 720         lr->lr_async_windows = bk->lb_async_windows;
 721         if (pos->lp_oit_cookie <= 1)
 722                 lr->lr_param |= LPF_RESET;
 723
 724         rc = lfsck_layout_master_notify_others(env, com, lr);
 725         if (rc != 0) {
 726                 CERROR("%s: fail to notify others for layout start: rc = %d\n",
 727                        lfsck_lfsck2name(lfsck), rc);
 728                 GOTO(fini, rc);
 729         }
 730
 731         spin_lock(&llmd->llmd_lock);
 732         thread_set_flags(athread, SVC_RUNNING);
 733         spin_unlock(&llmd->llmd_lock);
 734         wake_up_all(&mthread->t_ctl_waitq);
 735
 736         while (1) {
 737                 while (!list_empty(&llmd->llmd_req_list)) {
 738                         bool wakeup = false;
 739
 740                         l_wait_event(athread->t_ctl_waitq,
 741                                      bk->lb_async_windows == 0 ||
 742                                      atomic_read(&llmd->llmd_rpcs_in_flight) <
 743                                                 bk->lb_async_windows ||
 744                                      llmd->llmd_exit,
 745                                      &lwi);
 746
 747                         if (unlikely(llmd->llmd_exit))
 748                                 GOTO(cleanup1, rc = llmd->llmd_post_result);
 749
 750                         /* XXX: To be extended in other patch.
 751                          *
 752                          * Compare the OST side attribute with local attribute,
 753                          * and fix it if found inconsistency. */
 754
 755                         spin_lock(&llmd->llmd_lock);
 756                         llr = list_entry(llmd->llmd_req_list.next,
 757                                          struct lfsck_layout_req,
 758                                          llr_list);
 759                         list_del_init(&llr->llr_list);
 760                         if (bk->lb_async_windows != 0 &&
 761                             llmd->llmd_prefetched >= bk->lb_async_windows)
 762                                 wakeup = true;
 763
 764                         llmd->llmd_prefetched--;
 765                         spin_unlock(&llmd->llmd_lock);
 766                         if (wakeup)
 767                                 wake_up_all(&mthread->t_ctl_waitq);
 768
 769                         lfsck_layout_req_fini(env, llr);
 770                 }
 771
 772                 /* Wakeup the master engine if it is waiting in checkpoint. */
 773                 if (atomic_read(&llmd->llmd_rpcs_in_flight) == 0)
 774                         wake_up_all(&mthread->t_ctl_waitq);
 775
 776                 l_wait_event(athread->t_ctl_waitq,
 777                              !lfsck_layout_req_empty(llmd) ||
 778                              llmd->llmd_exit ||
 779                              llmd->llmd_to_post ||
 780                              llmd->llmd_to_double_scan,
 781                              &lwi);
 782
 783                 if (unlikely(llmd->llmd_exit))
 784                         GOTO(cleanup1, rc = llmd->llmd_post_result);
 785
 786                 if (!list_empty(&llmd->llmd_req_list))
 787                         continue;
 788
 789                 if (llmd->llmd_to_post) {
 790                         llmd->llmd_to_post = 0;
 791                         LASSERT(llmd->llmd_post_result > 0);
 792
 793                         memset(lr, 0, sizeof(*lr));
 794                         lr->lr_index = lfsck_dev_idx(lfsck->li_bottom);
 795                         lr->lr_active = LT_LAYOUT;
 796                         lr->lr_event = LE_PHASE1_DONE;
 797                         lr->lr_status = llmd->llmd_post_result;
 798                         rc = lfsck_layout_master_notify_others(env, com, lr);
 799                         if (rc != 0)
 800                                 CERROR("%s: failed to notify others "
 801                                        "for layout post: rc = %d\n",
 802                                        lfsck_lfsck2name(lfsck), rc);
 803
 804                         /* Wakeup the master engine to go ahead. */
 805                         wake_up_all(&mthread->t_ctl_waitq);
 806                 }
 807
 808                 if (llmd->llmd_to_double_scan) {
 809                         llmd->llmd_to_double_scan = 0;
 810                         atomic_inc(&lfsck->li_double_scan_count);
 811                         llmd->llmd_in_double_scan = 1;
 812                         wake_up_all(&mthread->t_ctl_waitq);
 813
 814                         while (llmd->llmd_in_double_scan) {
 815                                 rc = lfsck_layout_master_query_others(env, com);
 816                                 if (lfsck_layout_master_to_orphan(llmd))
 817                                         goto orphan;
 818
 819                                 if (rc < 0)
 820                                         GOTO(cleanup2, rc);
 821
 822                                 /* Pull LFSCK status on related targets once
 823                                  * per 30 seconds if we are not notified. */
 824                                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(30),
 825                                                            cfs_time_seconds(1),
 826                                                            NULL, NULL);
 827                                 rc = l_wait_event(athread->t_ctl_waitq,
 828                                         lfsck_layout_master_to_orphan(llmd) ||
 829                                         llmd->llmd_exit ||
 830                                         !thread_is_running(mthread),
 831                                         &lwi);
 832
 833                                 if (unlikely(llmd->llmd_exit ||
 834                                              !thread_is_running(mthread)))
 835                                         GOTO(cleanup2, rc = 0);
 836
 837                                 if (rc == -ETIMEDOUT)
 838                                         continue;
 839
 840                                 if (rc < 0)
 841                                         GOTO(cleanup2, rc);
 842
 843 orphan:
 844                                 /* XXX: real double scan for ost orphans. */
 845
 846                                 GOTO(cleanup2, rc = 1);
 847                         }
 848                 }
 849         }
 850
 851 cleanup1:
 852         /* Cleanup the unfinished requests. */
 853         spin_lock(&llmd->llmd_lock);
 854         while (!list_empty(&llmd->llmd_req_list)) {
 855                 llr = list_entry(llmd->llmd_req_list.next,
 856                                  struct lfsck_layout_req,
 857                                  llr_list);
 858                 list_del_init(&llr->llr_list);
 859                 llmd->llmd_prefetched--;
 860                 spin_unlock(&llmd->llmd_lock);
 861                 lfsck_layout_req_fini(env, llr);
 862                 spin_lock(&llmd->llmd_lock);
 863         }
 864         spin_unlock(&llmd->llmd_lock);
 865
 866         LASSERTF(llmd->llmd_prefetched == 0, "unmatched prefeteched objs %d\n",
 867                  llmd->llmd_prefetched);
 868
 869         l_wait_event(athread->t_ctl_waitq,
 870                      atomic_read(&llmd->llmd_rpcs_in_flight) == 0,
 871                      &lwi);
 872
 873 cleanup2:
 874         memset(lr, 0, sizeof(*lr));
 875         lr->lr_index = lfsck_dev_idx(lfsck->li_bottom);
 876         lr->lr_active = LT_LAYOUT;
 877         if (rc > 0) {
 878                 lr->lr_event = LE_PHASE2_DONE;
 879                 lr->lr_status = rc;
 880         } else if (rc == 0) {
 881                 lr->lr_event = LE_STOP;
 882                 if (lfsck->li_paused)
 883                         lr->lr_status = LS_CO_PAUSED;
 884                 else
 885                         lr->lr_status = LS_CO_STOPPED;
 886         } else {
 887                 lr->lr_event = LE_STOP;
 888                 lr->lr_status = LS_CO_FAILED;
 889         }
 890
 891         rc1 = lfsck_layout_master_notify_others(env, com, lr);
 892         if (rc1 != 0) {
 893                 CERROR("%s: failed to notify others for layout quit: rc = %d\n",
 894                        lfsck_lfsck2name(lfsck), rc1);
 895                 rc = rc1;
 896         }
 897
 898         /* Under force exit case, some requests may be just freed without
 899          * verification, those objects should be re-handled when next run.
 900          * So not update the on-disk tracing file under such case. */
 901         if (!llmd->llmd_exit)
 902                 rc1 = lfsck_layout_double_scan_result(env, com, rc);
 903
 904 fini:
 905         if (llmd->llmd_in_double_scan)
 906                 atomic_dec(&lfsck->li_double_scan_count);
 907
 908         spin_lock(&llmd->llmd_lock);
 909         llmd->llmd_assistant_status = (rc1 != 0 ? rc1 : rc);
 910         thread_set_flags(athread, SVC_STOPPED);
 911         wake_up_all(&mthread->t_ctl_waitq);
 912         spin_unlock(&llmd->llmd_lock);
 913         lfsck_thread_args_fini(lta);
 914
 915         return rc;
 916 }
 917
 918 /* layout APIs */
 919
 920 static int lfsck_layout_reset(const struct lu_env *env,
 921                               struct lfsck_component *com, bool init)
 922 {
 923         struct lfsck_layout     *lo    = com->lc_file_ram;
 924         int                      rc;
 925
 926         down_write(&com->lc_sem);
 927         if (init) {
 928                 memset(lo, 0, com->lc_file_size);
 929         } else {
 930                 __u32 count = lo->ll_success_count;
 931                 __u64 last_time = lo->ll_time_last_complete;
 932
 933                 memset(lo, 0, com->lc_file_size);
 934                 lo->ll_success_count = count;
 935                 lo->ll_time_last_complete = last_time;
 936         }
 937
 938         lo->ll_magic = LFSCK_LAYOUT_MAGIC;
 939         lo->ll_status = LS_INIT;
 940
 941         rc = lfsck_layout_store(env, com);
 942         up_write(&com->lc_sem);
 943
 944         return rc;
 945 }
 946
 947 static void lfsck_layout_fail(const struct lu_env *env,
 948                               struct lfsck_component *com, bool new_checked)
 949 {
 950         struct lfsck_layout *lo = com->lc_file_ram;
 951
 952         down_write(&com->lc_sem);
 953         if (new_checked)
 954                 com->lc_new_checked++;
 955         lo->ll_objs_failed_phase1++;
 956         if (lo->ll_pos_first_inconsistent == 0) {
 957                 struct lfsck_instance *lfsck = com->lc_lfsck;
 958
 959                 lo->ll_pos_first_inconsistent =
 960                         lfsck->li_obj_oit->do_index_ops->dio_it.store(env,
 961                                                         lfsck->li_di_oit);
 962         }
 963         up_write(&com->lc_sem);
 964 }
 965
 966 static int lfsck_layout_master_checkpoint(const struct lu_env *env,
 967                                           struct lfsck_component *com, bool init)
 968 {
 969         struct lfsck_instance           *lfsck   = com->lc_lfsck;
 970         struct lfsck_layout             *lo      = com->lc_file_ram;
 971         struct lfsck_layout_master_data *llmd    = com->lc_data;
 972         struct ptlrpc_thread            *mthread = &lfsck->li_thread;
 973         struct ptlrpc_thread            *athread = &llmd->llmd_thread;
 974         struct l_wait_info               lwi     = { 0 };
 975         int                              rc;
 976
 977         if (com->lc_new_checked == 0 && !init)
 978                 return 0;
 979
 980         l_wait_event(mthread->t_ctl_waitq,
 981                      (list_empty(&llmd->llmd_req_list) &&
 982                       atomic_read(&llmd->llmd_rpcs_in_flight) == 0) ||
 983                      !thread_is_running(mthread) ||
 984                      thread_is_stopped(athread),
 985                      &lwi);
 986
 987         if (!thread_is_running(mthread) || thread_is_stopped(athread))
 988                 return 0;
 989
 990         down_write(&com->lc_sem);
 991         if (init) {
 992                 lo->ll_pos_latest_start = lfsck->li_pos_current.lp_oit_cookie;
 993         } else {
 994                 lo->ll_pos_last_checkpoint =
 995                                         lfsck->li_pos_current.lp_oit_cookie;
 996                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
 997                                 HALF_SEC - lfsck->li_time_last_checkpoint);
 998                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
 999                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
1000                 com->lc_new_checked = 0;
1001         }
1002
1003         rc = lfsck_layout_store(env, com);
1004         up_write(&com->lc_sem);
1005
1006         return rc;
1007 }
1008
1009 static int lfsck_layout_slave_checkpoint(const struct lu_env *env,
1010                                          struct lfsck_component *com, bool init)
1011 {
1012         struct lfsck_instance   *lfsck = com->lc_lfsck;
1013         struct lfsck_layout     *lo    = com->lc_file_ram;
1014         int                      rc;
1015
1016         if (com->lc_new_checked == 0 && !init)
1017                 return 0;
1018
1019         down_write(&com->lc_sem);
1020
1021         if (init) {
1022                 lo->ll_pos_latest_start = lfsck->li_pos_current.lp_oit_cookie;
1023         } else {
1024                 lo->ll_pos_last_checkpoint =
1025                                         lfsck->li_pos_current.lp_oit_cookie;
1026                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
1027                                 HALF_SEC - lfsck->li_time_last_checkpoint);
1028                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
1029                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
1030                 com->lc_new_checked = 0;
1031         }
1032
1033         rc = lfsck_layout_store(env, com);
1034
1035         up_write(&com->lc_sem);
1036
1037         return rc;
1038 }
1039
1040 static int lfsck_layout_slave_prep(const struct lu_env *env,
1041                                    struct lfsck_component *com)
1042 {
1043         struct lfsck_instance   *lfsck  = com->lc_lfsck;
1044         struct lfsck_layout     *lo     = com->lc_file_ram;
1045         struct lfsck_position   *pos    = &com->lc_pos_start;
1046
1047         /* XXX: For a new scanning, generate OST-objects
1048          *      bitmap for orphan detection. */
1049
1050         fid_zero(&pos->lp_dir_parent);
1051         pos->lp_dir_cookie = 0;
1052         if (lo->ll_status == LS_COMPLETED ||
1053             lo->ll_status == LS_PARTIAL) {
1054                 int rc;
1055
1056                 rc = lfsck_layout_reset(env, com, false);
1057                 if (rc != 0)
1058                         return rc;
1059         }
1060
1061         down_write(&com->lc_sem);
1062
1063         lo->ll_time_latest_start = cfs_time_current_sec();
1064
1065         spin_lock(&lfsck->li_lock);
1066         if (lo->ll_flags & LF_SCANNED_ONCE) {
1067                 if (!lfsck->li_drop_dryrun ||
1068                     lo->ll_pos_first_inconsistent == 0) {
1069                         lo->ll_status = LS_SCANNING_PHASE2;
1070                         list_del_init(&com->lc_link);
1071                         list_add_tail(&com->lc_link,
1072                                       &lfsck->li_list_double_scan);
1073                         pos->lp_oit_cookie = 0;
1074                 } else {
1075                         int i;
1076
1077                         lo->ll_status = LS_SCANNING_PHASE1;
1078                         lo->ll_run_time_phase1 = 0;
1079                         lo->ll_run_time_phase2 = 0;
1080                         lo->ll_objs_checked_phase1 = 0;
1081                         lo->ll_objs_checked_phase2 = 0;
1082                         lo->ll_objs_failed_phase1 = 0;
1083                         lo->ll_objs_failed_phase2 = 0;
1084                         for (i = 0; i < LLIT_MAX; i++)
1085                                 lo->ll_objs_repaired[i] = 0;
1086
1087                         pos->lp_oit_cookie = lo->ll_pos_first_inconsistent;
1088                 }
1089         } else {
1090                 lo->ll_status = LS_SCANNING_PHASE1;
1091                 if (!lfsck->li_drop_dryrun ||
1092                     lo->ll_pos_first_inconsistent == 0)
1093                         pos->lp_oit_cookie = lo->ll_pos_last_checkpoint + 1;
1094                 else
1095                         pos->lp_oit_cookie = lo->ll_pos_first_inconsistent;
1096         }
1097         spin_unlock(&lfsck->li_lock);
1098
1099         up_write(&com->lc_sem);
1100
1101         return 0;
1102 }
1103
1104 static int lfsck_layout_master_prep(const struct lu_env *env,
1105                                     struct lfsck_component *com)
1106 {
1107         struct lfsck_instance           *lfsck   = com->lc_lfsck;
1108         struct lfsck_layout_master_data *llmd    = com->lc_data;
1109         struct ptlrpc_thread            *mthread = &lfsck->li_thread;
1110         struct ptlrpc_thread            *athread = &llmd->llmd_thread;
1111         struct lfsck_thread_args        *lta;
1112         long                             rc;
1113         ENTRY;
1114
1115         rc = lfsck_layout_slave_prep(env, com);
1116         if (rc != 0)
1117                 RETURN(rc);
1118
1119         llmd->llmd_assistant_status = 0;
1120         llmd->llmd_post_result = 0;
1121         llmd->llmd_to_post = 0;
1122         llmd->llmd_to_double_scan = 0;
1123         llmd->llmd_in_double_scan = 0;
1124         llmd->llmd_exit = 0;
1125         thread_set_flags(athread, 0);
1126
1127         lta = lfsck_thread_args_init(lfsck, com);
1128         if (IS_ERR(lta))
1129                 RETURN(PTR_ERR(lta));
1130
1131         rc = PTR_ERR(kthread_run(lfsck_layout_assistant, lta, "lfsck_layout"));
1132         if (IS_ERR_VALUE(rc)) {
1133                 CERROR("%s: Cannot start LFSCK layout assistant thread: "
1134                        "rc = %ld\n", lfsck_lfsck2name(lfsck), rc);
1135                 lfsck_thread_args_fini(lta);
1136         } else {
1137                 struct l_wait_info lwi = { 0 };
1138
1139                 l_wait_event(mthread->t_ctl_waitq,
1140                              thread_is_running(athread) ||
1141                              thread_is_stopped(athread),
1142                              &lwi);
1143                 if (unlikely(!thread_is_running(athread)))
1144                         rc = llmd->llmd_assistant_status;
1145                 else
1146                         rc = 0;
1147         }
1148
1149         RETURN(rc);
1150 }
1151
1152 static int lfsck_layout_master_exec_oit(const struct lu_env *env,
1153                                         struct lfsck_component *com,
1154                                         struct dt_object *obj)
1155 {
1156         /* XXX: To be implemented in other patches.
1157          *
1158          * For the given object, read its layout EA locally. For each stripe,
1159          * pre-fetch the OST-object's attribute and generate an structure
1160          * lfsck_layout_req on the list ::llmd_req_list.
1161          *
1162          * For each request on the ::llmd_req_list, the lfsck_layout_assistant
1163          * thread will compare the OST side attribute with local attribute,
1164          * if inconsistent, then repair it.
1165          *
1166          * All above processing is async mode with pipeline. */
1167
1168         return 0;
1169 }
1170
1171 static int lfsck_layout_slave_exec_oit(const struct lu_env *env,
1172                                        struct lfsck_component *com,
1173                                        struct dt_object *obj)
1174 {
1175         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1176         struct lfsck_layout             *lo     = com->lc_file_ram;
1177         const struct lu_fid             *fid    = lfsck_dto2fid(obj);
1178         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
1179         struct lfsck_layout_seq         *lls;
1180         __u64                            seq;
1181         __u64                            oid;
1182         int                              rc;
1183         ENTRY;
1184
1185         /* XXX: Update OST-objects bitmap for orphan detection. */
1186
1187         LASSERT(llsd != NULL);
1188
1189         down_write(&com->lc_sem);
1190         if (fid_is_idif(fid))
1191                 seq = 0;
1192         else if (!fid_is_norm(fid) ||
1193                  !fid_is_for_ostobj(env, lfsck->li_next, obj, fid))
1194                 GOTO(unlock, rc = 0);
1195         else
1196                 seq = fid_seq(fid);
1197         com->lc_new_checked++;
1198
1199         lls = lfsck_layout_seq_lookup(llsd, seq);
1200         if (lls == NULL) {
1201                 OBD_ALLOC_PTR(lls);
1202                 if (unlikely(lls == NULL))
1203                         GOTO(unlock, rc = -ENOMEM);
1204
1205                 INIT_LIST_HEAD(&lls->lls_list);
1206                 lls->lls_seq = seq;
1207                 rc = lfsck_layout_lastid_load(env, com, lls);
1208                 if (rc != 0) {
1209                         lo->ll_objs_failed_phase1++;
1210                         OBD_FREE_PTR(lls);
1211                         GOTO(unlock, rc);
1212                 }
1213
1214                 lfsck_layout_seq_insert(llsd, lls);
1215         }
1216
1217         if (unlikely(fid_is_last_id(fid)))
1218                 GOTO(unlock, rc = 0);
1219
1220         oid = fid_oid(fid);
1221         if (oid > lls->lls_lastid_known)
1222                 lls->lls_lastid_known = oid;
1223
1224         if (oid > lls->lls_lastid) {
1225                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
1226                         /* OFD may create new objects during LFSCK scanning. */
1227                         rc = lfsck_layout_lastid_reload(env, com, lls);
1228                         if (unlikely(rc != 0))
1229                                 CWARN("%s: failed to reload LAST_ID for "LPX64
1230                                       ": rc = %d\n",
1231                                       lfsck_lfsck2name(com->lc_lfsck),
1232                                       lls->lls_seq, rc);
1233                         if (oid <= lls->lls_lastid)
1234                                 GOTO(unlock, rc = 0);
1235
1236                         LASSERT(lfsck->li_out_notify != NULL);
1237
1238                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1239                                              LE_LASTID_REBUILDING);
1240                         lo->ll_flags |= LF_CRASHED_LASTID;
1241                 }
1242
1243                 lls->lls_lastid = oid;
1244                 lls->lls_dirty = 1;
1245         }
1246
1247         GOTO(unlock, rc = 0);
1248
1249 unlock:
1250         up_write(&com->lc_sem);
1251
1252         return rc;
1253 }
1254
1255 static int lfsck_layout_exec_dir(const struct lu_env *env,
1256                                  struct lfsck_component *com,
1257                                  struct dt_object *obj,
1258                                  struct lu_dirent *ent)
1259 {
1260         return 0;
1261 }
1262
1263 static int lfsck_layout_master_post(const struct lu_env *env,
1264                                     struct lfsck_component *com,
1265                                     int result, bool init)
1266 {
1267         struct lfsck_instance           *lfsck   = com->lc_lfsck;
1268         struct lfsck_layout             *lo      = com->lc_file_ram;
1269         struct lfsck_layout_master_data *llmd    = com->lc_data;
1270         struct ptlrpc_thread            *mthread = &lfsck->li_thread;
1271         struct ptlrpc_thread            *athread = &llmd->llmd_thread;
1272         struct l_wait_info               lwi     = { 0 };
1273         int                              rc;
1274         ENTRY;
1275
1276
1277         llmd->llmd_post_result = result;
1278         llmd->llmd_to_post = 1;
1279         if (llmd->llmd_post_result <= 0)
1280                 llmd->llmd_exit = 1;
1281
1282         wake_up_all(&athread->t_ctl_waitq);
1283         l_wait_event(mthread->t_ctl_waitq,
1284                      (result > 0 && list_empty(&llmd->llmd_req_list) &&
1285                       atomic_read(&llmd->llmd_rpcs_in_flight) == 0) ||
1286                      thread_is_stopped(athread),
1287                      &lwi);
1288
1289         if (llmd->llmd_assistant_status < 0)
1290                 result = llmd->llmd_assistant_status;
1291
1292         down_write(&com->lc_sem);
1293         spin_lock(&lfsck->li_lock);
1294         /* When LFSCK failed, there may be some prefetched objects those are
1295          * not been processed yet, we do not know the exactly position, then
1296          * just restart from last check-point next time. */
1297         if (!init && !llmd->llmd_exit)
1298                 lo->ll_pos_last_checkpoint =
1299                                         lfsck->li_pos_current.lp_oit_cookie;
1300
1301         if (result > 0) {
1302                 lo->ll_status = LS_SCANNING_PHASE2;
1303                 lo->ll_flags |= LF_SCANNED_ONCE;
1304                 lo->ll_flags &= ~LF_UPGRADE;
1305                 list_del_init(&com->lc_link);
1306                 list_add_tail(&com->lc_link, &lfsck->li_list_double_scan);
1307         } else if (result == 0) {
1308                 if (lfsck->li_paused) {
1309                         lo->ll_status = LS_PAUSED;
1310                 } else {
1311                         lo->ll_status = LS_STOPPED;
1312                         list_del_init(&com->lc_link);
1313                         list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1314                 }
1315         } else {
1316                 lo->ll_status = LS_FAILED;
1317                 list_del_init(&com->lc_link);
1318                 list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1319         }
1320         spin_unlock(&lfsck->li_lock);
1321
1322         if (!init) {
1323                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
1324                                 HALF_SEC - lfsck->li_time_last_checkpoint);
1325                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
1326                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
1327                 com->lc_new_checked = 0;
1328         }
1329
1330         rc = lfsck_layout_store(env, com);
1331         up_write(&com->lc_sem);
1332
1333         RETURN(rc);
1334 }
1335
1336 static int lfsck_layout_slave_post(const struct lu_env *env,
1337                                    struct lfsck_component *com,
1338                                    int result, bool init)
1339 {
1340         struct lfsck_instance   *lfsck = com->lc_lfsck;
1341         struct lfsck_layout     *lo    = com->lc_file_ram;
1342         int                      rc;
1343         bool                     done  = false;
1344
1345         rc = lfsck_layout_lastid_store(env, com);
1346         if (rc != 0)
1347                 result = rc;
1348
1349         LASSERT(lfsck->li_out_notify != NULL);
1350
1351         down_write(&com->lc_sem);
1352
1353         spin_lock(&lfsck->li_lock);
1354         if (!init)
1355                 lo->ll_pos_last_checkpoint =
1356                                         lfsck->li_pos_current.lp_oit_cookie;
1357         if (result > 0) {
1358                 lo->ll_status = LS_SCANNING_PHASE2;
1359                 lo->ll_flags |= LF_SCANNED_ONCE;
1360                 if (lo->ll_flags & LF_CRASHED_LASTID) {
1361                         done = true;
1362                         lo->ll_flags &= ~LF_CRASHED_LASTID;
1363                 }
1364                 lo->ll_flags &= ~LF_UPGRADE;
1365                 list_del_init(&com->lc_link);
1366                 list_add_tail(&com->lc_link, &lfsck->li_list_double_scan);
1367         } else if (result == 0) {
1368                 if (lfsck->li_paused) {
1369                         lo->ll_status = LS_PAUSED;
1370                 } else {
1371                         lo->ll_status = LS_STOPPED;
1372                         list_del_init(&com->lc_link);
1373                         list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1374                 }
1375         } else {
1376                 lo->ll_status = LS_FAILED;
1377                 list_del_init(&com->lc_link);
1378                 list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1379         }
1380         spin_unlock(&lfsck->li_lock);
1381
1382         if (done)
1383                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1384                                      LE_LASTID_REBUILT);
1385
1386         if (!init) {
1387                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
1388                                 HALF_SEC - lfsck->li_time_last_checkpoint);
1389                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
1390                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
1391                 com->lc_new_checked = 0;
1392         }
1393
1394         rc = lfsck_layout_store(env, com);
1395
1396         up_write(&com->lc_sem);
1397
1398         return rc;
1399 }
1400
1401 static int lfsck_layout_dump(const struct lu_env *env,
1402                              struct lfsck_component *com, char *buf, int len)
1403 {
1404         struct lfsck_instance   *lfsck = com->lc_lfsck;
1405         struct lfsck_bookmark   *bk    = &lfsck->li_bookmark_ram;
1406         struct lfsck_layout     *lo    = com->lc_file_ram;
1407         int                      save  = len;
1408         int                      ret   = -ENOSPC;
1409         int                      rc;
1410
1411         down_read(&com->lc_sem);
1412         rc = snprintf(buf, len,
1413                       "name: lfsck_layout\n"
1414                       "magic: %#x\n"
1415                       "version: %d\n"
1416                       "status: %s\n",
1417                       lo->ll_magic,
1418                       bk->lb_version,
1419                       lfsck_status2names(lo->ll_status));
1420         if (rc <= 0)
1421                 goto out;
1422
1423         buf += rc;
1424         len -= rc;
1425         rc = lfsck_bits_dump(&buf, &len, lo->ll_flags, lfsck_flags_names,
1426                              "flags");
1427         if (rc < 0)
1428                 goto out;
1429
1430         rc = lfsck_bits_dump(&buf, &len, bk->lb_param, lfsck_param_names,
1431                              "param");
1432         if (rc < 0)
1433                 goto out;
1434
1435         rc = lfsck_time_dump(&buf, &len, lo->ll_time_last_complete,
1436                              "time_since_last_completed");
1437         if (rc < 0)
1438                 goto out;
1439
1440         rc = lfsck_time_dump(&buf, &len, lo->ll_time_latest_start,
1441                              "time_since_latest_start");
1442         if (rc < 0)
1443                 goto out;
1444
1445         rc = lfsck_time_dump(&buf, &len, lo->ll_time_last_checkpoint,
1446                              "time_since_last_checkpoint");
1447         if (rc < 0)
1448                 goto out;
1449
1450         rc = snprintf(buf, len,
1451                       "latest_start_position: "LPU64"\n"
1452                       "last_checkpoint_position: "LPU64"\n"
1453                       "first_failure_position: "LPU64"\n",
1454                       lo->ll_pos_latest_start,
1455                       lo->ll_pos_last_checkpoint,
1456                       lo->ll_pos_first_inconsistent);
1457         if (rc <= 0)
1458                 goto out;
1459
1460         buf += rc;
1461         len -= rc;
1462
1463         rc = snprintf(buf, len,
1464                       "success_count: %u\n"
1465                       "repaired_dangling: "LPU64"\n"
1466                       "repaired_unmatched_pair: "LPU64"\n"
1467                       "repaired_multiple_referenced: "LPU64"\n"
1468                       "repaired_orphan: "LPU64"\n"
1469                       "repaired_inconsistent_owner: "LPU64"\n"
1470                       "repaired_others: "LPU64"\n"
1471                       "skipped: "LPU64"\n"
1472                       "failed_phase1: "LPU64"\n"
1473                       "failed_phase2: "LPU64"\n",
1474                       lo->ll_success_count,
1475                       lo->ll_objs_repaired[LLIT_DANGLING - 1],
1476                       lo->ll_objs_repaired[LLIT_UNMATCHED_PAIR - 1],
1477                       lo->ll_objs_repaired[LLIT_MULTIPLE_REFERENCED - 1],
1478                       lo->ll_objs_repaired[LLIT_ORPHAN - 1],
1479                       lo->ll_objs_repaired[LLIT_INCONSISTENT_OWNER - 1],
1480                       lo->ll_objs_repaired[LLIT_OTHERS - 1],
1481                       lo->ll_objs_skipped,
1482                       lo->ll_objs_failed_phase1,
1483                       lo->ll_objs_failed_phase2);
1484         if (rc <= 0)
1485                 goto out;
1486
1487         buf += rc;
1488         len -= rc;
1489
1490         if (lo->ll_status == LS_SCANNING_PHASE1) {
1491                 __u64 pos;
1492                 const struct dt_it_ops *iops;
1493                 cfs_duration_t duration = cfs_time_current() -
1494                                           lfsck->li_time_last_checkpoint;
1495                 __u64 checked = lo->ll_objs_checked_phase1 + com->lc_new_checked;
1496                 __u64 speed = checked;
1497                 __u64 new_checked = com->lc_new_checked * HZ;
1498                 __u32 rtime = lo->ll_run_time_phase1 +
1499                               cfs_duration_sec(duration + HALF_SEC);
1500
1501                 if (duration != 0)
1502                         do_div(new_checked, duration);
1503                 if (rtime != 0)
1504                         do_div(speed, rtime);
1505                 rc = snprintf(buf, len,
1506                               "checked_phase1: "LPU64"\n"
1507                               "checked_phase2: "LPU64"\n"
1508                               "run_time_phase1: %u seconds\n"
1509                               "run_time_phase2: %u seconds\n"
1510                               "average_speed_phase1: "LPU64" items/sec\n"
1511                               "average_speed_phase2: N/A\n"
1512                               "real-time_speed_phase1: "LPU64" items/sec\n"
1513                               "real-time_speed_phase2: N/A\n",
1514                               checked,
1515                               lo->ll_objs_checked_phase2,
1516                               rtime,
1517                               lo->ll_run_time_phase2,
1518                               speed,
1519                               new_checked);
1520                 if (rc <= 0)
1521                         goto out;
1522
1523                 buf += rc;
1524                 len -= rc;
1525
1526                 LASSERT(lfsck->li_di_oit != NULL);
1527
1528                 iops = &lfsck->li_obj_oit->do_index_ops->dio_it;
1529
1530                 /* The low layer otable-based iteration position may NOT
1531                  * exactly match the layout-based directory traversal
1532                  * cookie. Generally, it is not a serious issue. But the
1533                  * caller should NOT make assumption on that. */
1534                 pos = iops->store(env, lfsck->li_di_oit);
1535                 if (!lfsck->li_current_oit_processed)
1536                         pos--;
1537                 rc = snprintf(buf, len, "current_position: "LPU64"\n", pos);
1538                 if (rc <= 0)
1539                         goto out;
1540
1541                 buf += rc;
1542                 len -= rc;
1543         } else {
1544                 /* XXX: LS_SCANNING_PHASE2 will be handled in the future. */
1545                 __u64 speed1 = lo->ll_objs_checked_phase1;
1546                 __u64 speed2 = lo->ll_objs_checked_phase2;
1547
1548                 if (lo->ll_run_time_phase1 != 0)
1549                         do_div(speed1, lo->ll_run_time_phase1);
1550                 if (lo->ll_run_time_phase2 != 0)
1551                         do_div(speed2, lo->ll_run_time_phase2);
1552                 rc = snprintf(buf, len,
1553                               "checked_phase1: "LPU64"\n"
1554                               "checked_phase2: "LPU64"\n"
1555                               "run_time_phase1: %u seconds\n"
1556                               "run_time_phase2: %u seconds\n"
1557                               "average_speed_phase1: "LPU64" items/sec\n"
1558                               "average_speed_phase2: "LPU64" objs/sec\n"
1559                               "real-time_speed_phase1: N/A\n"
1560                               "real-time_speed_phase2: N/A\n"
1561                               "current_position: N/A\n",
1562                               lo->ll_objs_checked_phase1,
1563                               lo->ll_objs_checked_phase2,
1564                               lo->ll_run_time_phase1,
1565                               lo->ll_run_time_phase2,
1566                               speed1,
1567                               speed2);
1568                 if (rc <= 0)
1569                         goto out;
1570
1571                 buf += rc;
1572                 len -= rc;
1573         }
1574         ret = save - len;
1575
1576 out:
1577         up_read(&com->lc_sem);
1578
1579         return ret;
1580 }
1581
1582 static int lfsck_layout_master_double_scan(const struct lu_env *env,
1583                                            struct lfsck_component *com)
1584 {
1585         struct lfsck_layout_master_data *llmd    = com->lc_data;
1586         struct ptlrpc_thread            *mthread = &com->lc_lfsck->li_thread;
1587         struct ptlrpc_thread            *athread = &llmd->llmd_thread;
1588         struct lfsck_layout             *lo      = com->lc_file_ram;
1589         struct l_wait_info               lwi     = { 0 };
1590
1591         if (unlikely(lo->ll_status != LS_SCANNING_PHASE2))
1592                 return 0;
1593
1594         llmd->llmd_to_double_scan = 1;
1595         wake_up_all(&athread->t_ctl_waitq);
1596         l_wait_event(mthread->t_ctl_waitq,
1597                      llmd->llmd_in_double_scan ||
1598                      thread_is_stopped(athread),
1599                      &lwi);
1600         if (llmd->llmd_assistant_status < 0)
1601                 return llmd->llmd_assistant_status;
1602
1603         return 0;
1604 }
1605
1606 static int lfsck_layout_slave_double_scan(const struct lu_env *env,
1607                                           struct lfsck_component *com)
1608 {
1609         struct lfsck_instance   *lfsck = com->lc_lfsck;
1610         struct lfsck_layout     *lo    = com->lc_file_ram;
1611         int                      rc    = 1;
1612
1613         if (unlikely(lo->ll_status != LS_SCANNING_PHASE2))
1614                 return 0;
1615
1616         atomic_inc(&lfsck->li_double_scan_count);
1617
1618         com->lc_new_checked = 0;
1619         com->lc_new_scanned = 0;
1620         com->lc_time_last_checkpoint = cfs_time_current();
1621         com->lc_time_next_checkpoint = com->lc_time_last_checkpoint +
1622                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
1623
1624         rc = lfsck_layout_double_scan_result(env, com, rc);
1625
1626         if (atomic_dec_and_test(&lfsck->li_double_scan_count))
1627                 wake_up_all(&lfsck->li_thread.t_ctl_waitq);
1628
1629         return rc;
1630 }
1631
1632 static void lfsck_layout_master_data_release(const struct lu_env *env,
1633                                              struct lfsck_component *com)
1634 {
1635         struct lfsck_layout_master_data *llmd = com->lc_data;
1636
1637         LASSERT(llmd != NULL);
1638         LASSERT(thread_is_init(&llmd->llmd_thread) ||
1639                 thread_is_stopped(&llmd->llmd_thread));
1640         LASSERT(list_empty(&llmd->llmd_req_list));
1641         LASSERT(atomic_read(&llmd->llmd_rpcs_in_flight) == 0);
1642
1643         com->lc_data = NULL;
1644         OBD_FREE_PTR(llmd);
1645 }
1646
1647 static void lfsck_layout_slave_data_release(const struct lu_env *env,
1648                                             struct lfsck_component *com)
1649 {
1650         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
1651         struct lfsck_layout_seq         *lls;
1652         struct lfsck_layout_seq         *next;
1653
1654         LASSERT(llsd != NULL);
1655
1656         com->lc_data = NULL;
1657
1658         list_for_each_entry_safe(lls, next, &llsd->llsd_seq_list,
1659                                      lls_list) {
1660                 list_del_init(&lls->lls_list);
1661                 lfsck_object_put(env, lls->lls_lastid_obj);
1662                 OBD_FREE_PTR(lls);
1663         }
1664
1665         OBD_FREE_PTR(llsd);
1666 }
1667
1668 static void lfsck_layout_master_quit(const struct lu_env *env,
1669                                      struct lfsck_component *com)
1670 {
1671         struct lfsck_layout_master_data *llmd    = com->lc_data;
1672         struct ptlrpc_thread            *mthread = &com->lc_lfsck->li_thread;
1673         struct ptlrpc_thread            *athread = &llmd->llmd_thread;
1674         struct l_wait_info               lwi     = { 0 };
1675
1676         llmd->llmd_exit = 1;
1677         wake_up_all(&athread->t_ctl_waitq);
1678         l_wait_event(mthread->t_ctl_waitq,
1679                      thread_is_init(athread) ||
1680                      thread_is_stopped(athread),
1681                      &lwi);
1682 }
1683
1684 static struct lfsck_operations lfsck_layout_master_ops = {
1685         .lfsck_reset            = lfsck_layout_reset,
1686         .lfsck_fail             = lfsck_layout_fail,
1687         .lfsck_checkpoint       = lfsck_layout_master_checkpoint,
1688         .lfsck_prep             = lfsck_layout_master_prep,
1689         .lfsck_exec_oit         = lfsck_layout_master_exec_oit,
1690         .lfsck_exec_dir         = lfsck_layout_exec_dir,
1691         .lfsck_post             = lfsck_layout_master_post,
1692         .lfsck_dump             = lfsck_layout_dump,
1693         .lfsck_double_scan      = lfsck_layout_master_double_scan,
1694         .lfsck_data_release     = lfsck_layout_master_data_release,
1695         .lfsck_quit             = lfsck_layout_master_quit,
1696 };
1697
1698 static struct lfsck_operations lfsck_layout_slave_ops = {
1699         .lfsck_reset            = lfsck_layout_reset,
1700         .lfsck_fail             = lfsck_layout_fail,
1701         .lfsck_checkpoint       = lfsck_layout_slave_checkpoint,
1702         .lfsck_prep             = lfsck_layout_slave_prep,
1703         .lfsck_exec_oit         = lfsck_layout_slave_exec_oit,
1704         .lfsck_exec_dir         = lfsck_layout_exec_dir,
1705         .lfsck_post             = lfsck_layout_slave_post,
1706         .lfsck_dump             = lfsck_layout_dump,
1707         .lfsck_double_scan      = lfsck_layout_slave_double_scan,
1708         .lfsck_data_release     = lfsck_layout_slave_data_release,
1709 };
1710
1711 int lfsck_layout_setup(const struct lu_env *env, struct lfsck_instance *lfsck)
1712 {
1713         struct lfsck_component  *com;
1714         struct lfsck_layout     *lo;
1715         struct dt_object        *root = NULL;
1716         struct dt_object        *obj;
1717         int                      rc;
1718         ENTRY;
1719
1720         OBD_ALLOC_PTR(com);
1721         if (com == NULL)
1722                 RETURN(-ENOMEM);
1723
1724         INIT_LIST_HEAD(&com->lc_link);
1725         INIT_LIST_HEAD(&com->lc_link_dir);
1726         init_rwsem(&com->lc_sem);
1727         atomic_set(&com->lc_ref, 1);
1728         com->lc_lfsck = lfsck;
1729         com->lc_type = LT_LAYOUT;
1730         if (lfsck->li_master) {
1731                 struct lfsck_layout_master_data *llmd;
1732
1733                 com->lc_ops = &lfsck_layout_master_ops;
1734                 OBD_ALLOC_PTR(llmd);
1735                 if (llmd == NULL)
1736                         GOTO(out, rc = -ENOMEM);
1737
1738                 INIT_LIST_HEAD(&llmd->llmd_req_list);
1739                 spin_lock_init(&llmd->llmd_lock);
1740                 init_waitqueue_head(&llmd->llmd_thread.t_ctl_waitq);
1741                 atomic_set(&llmd->llmd_rpcs_in_flight, 0);
1742                 com->lc_data = llmd;
1743         } else {
1744                 struct lfsck_layout_slave_data *llsd;
1745
1746                 com->lc_ops = &lfsck_layout_slave_ops;
1747                 OBD_ALLOC_PTR(llsd);
1748                 if (llsd == NULL)
1749                         GOTO(out, rc = -ENOMEM);
1750
1751                 INIT_LIST_HEAD(&llsd->llsd_seq_list);
1752                 com->lc_data = llsd;
1753         }
1754         com->lc_file_size = sizeof(*lo);
1755         OBD_ALLOC(com->lc_file_ram, com->lc_file_size);
1756         if (com->lc_file_ram == NULL)
1757                 GOTO(out, rc = -ENOMEM);
1758
1759         OBD_ALLOC(com->lc_file_disk, com->lc_file_size);
1760         if (com->lc_file_disk == NULL)
1761                 GOTO(out, rc = -ENOMEM);
1762
1763         root = dt_locate(env, lfsck->li_bottom, &lfsck->li_local_root_fid);
1764         if (IS_ERR(root))
1765                 GOTO(out, rc = PTR_ERR(root));
1766
1767         if (unlikely(!dt_try_as_dir(env, root)))
1768                 GOTO(out, rc = -ENOTDIR);
1769
1770         obj = local_file_find_or_create(env, lfsck->li_los, root,
1771                                         lfsck_layout_name,
1772                                         S_IFREG | S_IRUGO | S_IWUSR);
1773         if (IS_ERR(obj))
1774                 GOTO(out, rc = PTR_ERR(obj));
1775
1776         com->lc_obj = obj;
1777         rc = lfsck_layout_load(env, com);
1778         if (rc > 0)
1779                 rc = lfsck_layout_reset(env, com, true);
1780         else if (rc == -ENOENT)
1781                 rc = lfsck_layout_init(env, com);
1782
1783         if (rc != 0)
1784                 GOTO(out, rc);
1785
1786         lo = com->lc_file_ram;
1787         switch (lo->ll_status) {
1788         case LS_INIT:
1789         case LS_COMPLETED:
1790         case LS_FAILED:
1791         case LS_STOPPED:
1792         case LS_PARTIAL:
1793                 spin_lock(&lfsck->li_lock);
1794                 list_add_tail(&com->lc_link, &lfsck->li_list_idle);
1795                 spin_unlock(&lfsck->li_lock);
1796                 break;
1797         default:
1798                 CERROR("%s: unknown lfsck_layout status: rc = %u\n",
1799                        lfsck_lfsck2name(lfsck), lo->ll_status);
1800                 /* fall through */
1801         case LS_SCANNING_PHASE1:
1802         case LS_SCANNING_PHASE2:
1803                 /* No need to store the status to disk right now.
1804                  * If the system crashed before the status stored,
1805                  * it will be loaded back when next time. */
1806                 lo->ll_status = LS_CRASHED;
1807                 lo->ll_flags |= LF_INCOMPLETE;
1808                 /* fall through */
1809         case LS_PAUSED:
1810         case LS_CRASHED:
1811         case LS_CO_FAILED:
1812         case LS_CO_STOPPED:
1813         case LS_CO_PAUSED:
1814                 spin_lock(&lfsck->li_lock);
1815                 list_add_tail(&com->lc_link, &lfsck->li_list_scan);
1816                 spin_unlock(&lfsck->li_lock);
1817                 break;
1818         }
1819
1820         if (lo->ll_flags & LF_CRASHED_LASTID) {
1821                 LASSERT(lfsck->li_out_notify != NULL);
1822
1823                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1824                                      LE_LASTID_REBUILDING);
1825         }
1826
1827         GOTO(out, rc = 0);
1828
1829 out:
1830         if (root != NULL && !IS_ERR(root))
1831                 lu_object_put(env, &root->do_lu);
1832
1833         if (rc != 0)
1834                 lfsck_component_cleanup(env, com);
1835
1836         return rc;
1837 }