lustre/lfsck/lfsck_lib.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) 2012, 2013, Intel Corporation.
  24  */
  25 /*
  26  * lustre/lfsck/lfsck_lib.c
  27  *
  28  * Author: Fan, Yong <fan.yong@intel.com>
  29  */
  30
  31 #define DEBUG_SUBSYSTEM S_LFSCK
  32
  33 #include <libcfs/list.h>
  34 #include <lu_object.h>
  35 #include <dt_object.h>
  36 #include <md_object.h>
  37 #include <lustre_fld.h>
  38 #include <lustre_lib.h>
  39 #include <lustre_net.h>
  40 #include <lustre_lfsck.h>
  41 #include <lustre/lustre_lfsck_user.h>
  42
  43 #include "lfsck_internal.h"
  44
  45 /* define lfsck thread key */
  46 LU_KEY_INIT(lfsck, struct lfsck_thread_info);
  47
  48 static void lfsck_key_fini(const struct lu_context *ctx,
  49                            struct lu_context_key *key, void *data)
  50 {
  51         struct lfsck_thread_info *info = data;
  52
  53         lu_buf_free(&info->lti_linkea_buf);
  54         lu_buf_free(&info->lti_big_buf);
  55         OBD_FREE_PTR(info);
  56 }
  57
  58 LU_CONTEXT_KEY_DEFINE(lfsck, LCT_MD_THREAD | LCT_DT_THREAD);
  59 LU_KEY_INIT_GENERIC(lfsck);
  60
  61 static CFS_LIST_HEAD(lfsck_instance_list);
  62 static struct list_head lfsck_ost_orphan_list;
  63 static struct list_head lfsck_mdt_orphan_list;
  64 static DEFINE_SPINLOCK(lfsck_instance_lock);
  65
  66 static const char *lfsck_status_names[] = {
  67         [LS_INIT]               = "init",
  68         [LS_SCANNING_PHASE1]    = "scanning-phase1",
  69         [LS_SCANNING_PHASE2]    = "scanning-phase2",
  70         [LS_COMPLETED]          = "completed",
  71         [LS_FAILED]             = "failed",
  72         [LS_STOPPED]            = "stopped",
  73         [LS_PAUSED]             = "paused",
  74         [LS_CRASHED]            = "crashed",
  75         [LS_PARTIAL]            = "partial",
  76         [LS_CO_FAILED]          = "co-failed",
  77         [LS_CO_STOPPED]         = "co-stopped",
  78         [LS_CO_PAUSED]          = "co-paused"
  79 };
  80
  81 const char *lfsck_flags_names[] = {
  82         "scanned-once",
  83         "inconsistent",
  84         "upgrade",
  85         "incomplete",
  86         "crashed_lastid",
  87         NULL
  88 };
  89
  90 const char *lfsck_param_names[] = {
  91         NULL,
  92         "failout",
  93         "dryrun",
  94         "all_targets",
  95         NULL
  96 };
  97
  98 const char *lfsck_status2names(enum lfsck_status status)
  99 {
 100         if (unlikely(status < 0 || status >= LS_MAX))
 101                 return "unknown";
 102
 103         return lfsck_status_names[status];
 104 }
 105
 106 static int lfsck_tgt_descs_init(struct lfsck_tgt_descs *ltds)
 107 {
 108         spin_lock_init(&ltds->ltd_lock);
 109         init_rwsem(&ltds->ltd_rw_sem);
 110         INIT_LIST_HEAD(&ltds->ltd_orphan);
 111         ltds->ltd_tgts_bitmap = CFS_ALLOCATE_BITMAP(BITS_PER_LONG);
 112         if (ltds->ltd_tgts_bitmap == NULL)
 113                 return -ENOMEM;
 114
 115         return 0;
 116 }
 117
 118 static void lfsck_tgt_descs_fini(struct lfsck_tgt_descs *ltds)
 119 {
 120         struct lfsck_tgt_desc   *ltd;
 121         struct lfsck_tgt_desc   *next;
 122         int                      idx;
 123
 124         down_write(&ltds->ltd_rw_sem);
 125
 126         list_for_each_entry_safe(ltd, next, &ltds->ltd_orphan,
 127                                  ltd_orphan_list) {
 128                 list_del_init(&ltd->ltd_orphan_list);
 129                 lfsck_tgt_put(ltd);
 130         }
 131
 132         if (unlikely(ltds->ltd_tgts_bitmap == NULL)) {
 133                 up_write(&ltds->ltd_rw_sem);
 134
 135                 return;
 136         }
 137
 138         cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
 139                 ltd = LTD_TGT(ltds, idx);
 140                 if (likely(ltd != NULL)) {
 141                         LASSERT(list_empty(&ltd->ltd_layout_list));
 142                         LASSERT(list_empty(&ltd->ltd_layout_phase_list));
 143
 144                         ltds->ltd_tgtnr--;
 145                         cfs_bitmap_clear(ltds->ltd_tgts_bitmap, idx);
 146                         LTD_TGT(ltds, idx) = NULL;
 147                         lfsck_tgt_put(ltd);
 148                 }
 149         }
 150
 151         LASSERTF(ltds->ltd_tgtnr == 0, "tgt count unmatched: %d\n",
 152                  ltds->ltd_tgtnr);
 153
 154         for (idx = 0; idx < TGT_PTRS; idx++) {
 155                 if (ltds->ltd_tgts_idx[idx] != NULL) {
 156                         OBD_FREE_PTR(ltds->ltd_tgts_idx[idx]);
 157                         ltds->ltd_tgts_idx[idx] = NULL;
 158                 }
 159         }
 160
 161         CFS_FREE_BITMAP(ltds->ltd_tgts_bitmap);
 162         ltds->ltd_tgts_bitmap = NULL;
 163         up_write(&ltds->ltd_rw_sem);
 164 }
 165
 166 static int __lfsck_add_target(const struct lu_env *env,
 167                               struct lfsck_instance *lfsck,
 168                               struct lfsck_tgt_desc *ltd,
 169                               bool for_ost, bool locked)
 170 {
 171         struct lfsck_tgt_descs *ltds;
 172         __u32                   index = ltd->ltd_index;
 173         int                     rc    = 0;
 174         ENTRY;
 175
 176         if (for_ost)
 177                 ltds = &lfsck->li_ost_descs;
 178         else
 179                 ltds = &lfsck->li_mdt_descs;
 180
 181         if (!locked)
 182                 down_write(&ltds->ltd_rw_sem);
 183
 184         LASSERT(ltds->ltd_tgts_bitmap != NULL);
 185
 186         if (index >= ltds->ltd_tgts_bitmap->size) {
 187                 __u32 newsize = max((__u32)ltds->ltd_tgts_bitmap->size,
 188                                     (__u32)BITS_PER_LONG);
 189                 cfs_bitmap_t *old_bitmap = ltds->ltd_tgts_bitmap;
 190                 cfs_bitmap_t *new_bitmap;
 191
 192                 while (newsize < index + 1)
 193                         newsize <<= 1;
 194
 195                 new_bitmap = CFS_ALLOCATE_BITMAP(newsize);
 196                 if (new_bitmap == NULL)
 197                         GOTO(unlock, rc = -ENOMEM);
 198
 199                 if (ltds->ltd_tgtnr > 0)
 200                         cfs_bitmap_copy(new_bitmap, old_bitmap);
 201                 ltds->ltd_tgts_bitmap = new_bitmap;
 202                 CFS_FREE_BITMAP(old_bitmap);
 203         }
 204
 205         if (cfs_bitmap_check(ltds->ltd_tgts_bitmap, index)) {
 206                 CERROR("%s: the device %s (%u) is registered already\n",
 207                        lfsck_lfsck2name(lfsck),
 208                        ltd->ltd_tgt->dd_lu_dev.ld_obd->obd_name, index);
 209                 GOTO(unlock, rc = -EEXIST);
 210         }
 211
 212         if (ltds->ltd_tgts_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
 213                 OBD_ALLOC_PTR(ltds->ltd_tgts_idx[index / TGT_PTRS_PER_BLOCK]);
 214                 if (ltds->ltd_tgts_idx[index / TGT_PTRS_PER_BLOCK] == NULL)
 215                         GOTO(unlock, rc = -ENOMEM);
 216         }
 217
 218         LTD_TGT(ltds, index) = ltd;
 219         cfs_bitmap_set(ltds->ltd_tgts_bitmap, index);
 220         ltds->ltd_tgtnr++;
 221
 222         GOTO(unlock, rc = 0);
 223
 224 unlock:
 225         if (!locked)
 226                 up_write(&ltds->ltd_rw_sem);
 227
 228         return rc;
 229 }
 230
 231 static int lfsck_add_target_from_orphan(const struct lu_env *env,
 232                                         struct lfsck_instance *lfsck)
 233 {
 234         struct lfsck_tgt_descs  *ltds    = &lfsck->li_ost_descs;
 235         struct lfsck_tgt_desc   *ltd;
 236         struct lfsck_tgt_desc   *next;
 237         struct list_head        *head    = &lfsck_ost_orphan_list;
 238         int                      rc;
 239         bool                     for_ost = true;
 240
 241 again:
 242         spin_lock(&lfsck_instance_lock);
 243         list_for_each_entry_safe(ltd, next, head, ltd_orphan_list) {
 244                 if (ltd->ltd_key == lfsck->li_bottom) {
 245                         list_del_init(&ltd->ltd_orphan_list);
 246                         list_add_tail(&ltd->ltd_orphan_list,
 247                                       &ltds->ltd_orphan);
 248                 }
 249         }
 250         spin_unlock(&lfsck_instance_lock);
 251
 252         down_write(&ltds->ltd_rw_sem);
 253         while (!list_empty(&ltds->ltd_orphan)) {
 254                 ltd = list_entry(ltds->ltd_orphan.next,
 255                                  struct lfsck_tgt_desc,
 256                                  ltd_orphan_list);
 257                 list_del_init(&ltd->ltd_orphan_list);
 258                 rc = __lfsck_add_target(env, lfsck, ltd, for_ost, true);
 259                 /* Do not hold the semaphore for too long time. */
 260                 up_write(&ltds->ltd_rw_sem);
 261                 if (rc != 0)
 262                         return rc;
 263
 264                 down_write(&ltds->ltd_rw_sem);
 265         }
 266         up_write(&ltds->ltd_rw_sem);
 267
 268         if (for_ost) {
 269                 ltds = &lfsck->li_mdt_descs;
 270                 head = &lfsck_mdt_orphan_list;
 271                 for_ost = false;
 272                 goto again;
 273         }
 274
 275         return 0;
 276 }
 277
 278 static inline struct lfsck_component *
 279 __lfsck_component_find(struct lfsck_instance *lfsck, __u16 type, cfs_list_t *list)
 280 {
 281         struct lfsck_component *com;
 282
 283         cfs_list_for_each_entry(com, list, lc_link) {
 284                 if (com->lc_type == type)
 285                         return com;
 286         }
 287         return NULL;
 288 }
 289
 290 struct lfsck_component *
 291 lfsck_component_find(struct lfsck_instance *lfsck, __u16 type)
 292 {
 293         struct lfsck_component *com;
 294
 295         spin_lock(&lfsck->li_lock);
 296         com = __lfsck_component_find(lfsck, type, &lfsck->li_list_scan);
 297         if (com != NULL)
 298                 goto unlock;
 299
 300         com = __lfsck_component_find(lfsck, type,
 301                                      &lfsck->li_list_double_scan);
 302         if (com != NULL)
 303                 goto unlock;
 304
 305         com = __lfsck_component_find(lfsck, type, &lfsck->li_list_idle);
 306
 307 unlock:
 308         if (com != NULL)
 309                 lfsck_component_get(com);
 310         spin_unlock(&lfsck->li_lock);
 311         return com;
 312 }
 313
 314 void lfsck_component_cleanup(const struct lu_env *env,
 315                              struct lfsck_component *com)
 316 {
 317         if (!cfs_list_empty(&com->lc_link))
 318                 cfs_list_del_init(&com->lc_link);
 319         if (!cfs_list_empty(&com->lc_link_dir))
 320                 cfs_list_del_init(&com->lc_link_dir);
 321
 322         lfsck_component_put(env, com);
 323 }
 324
 325 int lfsck_fid_alloc(const struct lu_env *env, struct lfsck_instance *lfsck,
 326                     struct lu_fid *fid, bool locked)
 327 {
 328         struct lfsck_bookmark   *bk = &lfsck->li_bookmark_ram;
 329         int                      rc = 0;
 330         ENTRY;
 331
 332         if (!locked)
 333                 mutex_lock(&lfsck->li_mutex);
 334
 335         rc = seq_client_alloc_fid(env, lfsck->li_seq, fid);
 336         if (rc >= 0) {
 337                 bk->lb_last_fid = *fid;
 338                 /* We do not care about whether the subsequent sub-operations
 339                  * failed or not. The worst case is that one FID is lost that
 340                  * is not a big issue for the LFSCK since it is relative rare
 341                  * for LFSCK create. */
 342                 rc = lfsck_bookmark_store(env, lfsck);
 343         }
 344
 345         if (!locked)
 346                 mutex_unlock(&lfsck->li_mutex);
 347
 348         RETURN(rc);
 349 }
 350
 351 static const char dot[] = ".";
 352 static const char dotdot[] = "..";
 353
 354 static int lfsck_create_lpf_local(const struct lu_env *env,
 355                                   struct lfsck_instance *lfsck,
 356                                   struct dt_object *parent,
 357                                   struct dt_object *child,
 358                                   struct lu_attr *la,
 359                                   struct dt_object_format *dof,
 360                                   const char *name)
 361 {
 362         struct dt_device        *dev    = lfsck->li_bottom;
 363         struct lfsck_bookmark   *bk     = &lfsck->li_bookmark_ram;
 364         struct dt_object        *bk_obj = lfsck->li_bookmark_obj;
 365         const struct lu_fid     *cfid   = lu_object_fid(&child->do_lu);
 366         struct thandle          *th     = NULL;
 367         loff_t                   pos    = 0;
 368         int                      len    = sizeof(struct lfsck_bookmark);
 369         int                      rc     = 0;
 370         ENTRY;
 371
 372         th = dt_trans_create(env, dev);
 373         if (IS_ERR(th))
 374                 RETURN(PTR_ERR(th));
 375
 376         /* 1a. create child */
 377         rc = dt_declare_create(env, child, la, NULL, dof, th);
 378         if (rc != 0)
 379                 GOTO(stop, rc);
 380
 381         /* 2a. increase child nlink */
 382         rc = dt_declare_ref_add(env, child, th);
 383         if (rc != 0)
 384                 GOTO(stop, rc);
 385
 386         /* 3a. insert name into parent dir */
 387         rc = dt_declare_insert(env, parent, (const struct dt_rec *)cfid,
 388                                (const struct dt_key *)name, th);
 389         if (rc != 0)
 390                 GOTO(stop, rc);
 391
 392         /* 4a. increase parent nlink */
 393         rc = dt_declare_ref_add(env, parent, th);
 394         if (rc != 0)
 395                 GOTO(stop, rc);
 396
 397         /* 5a. update bookmark */
 398         rc = dt_declare_record_write(env, bk_obj,
 399                                      lfsck_buf_get(env, bk, len), 0, th);
 400         if (rc != 0)
 401                 GOTO(stop, rc);
 402
 403         rc = dt_trans_start_local(env, dev, th);
 404         if (rc != 0)
 405                 GOTO(stop, rc);
 406
 407         dt_write_lock(env, child, 0);
 408         /* 1b.1 create child */
 409         rc = dt_create(env, child, la, NULL, dof, th);
 410         if (rc != 0)
 411                 GOTO(unlock, rc);
 412
 413         if (unlikely(!dt_try_as_dir(env, child)))
 414                 GOTO(unlock, rc = -ENOTDIR);
 415
 416         /* 1b.2 insert dot into child dir */
 417         rc = dt_insert(env, child, (const struct dt_rec *)cfid,
 418                        (const struct dt_key *)dot, th, BYPASS_CAPA, 1);
 419         if (rc != 0)
 420                 GOTO(unlock, rc);
 421
 422         /* 1b.3 insert dotdot into child dir */
 423         rc = dt_insert(env, child, (const struct dt_rec *)&LU_LPF_FID,
 424                        (const struct dt_key *)dotdot, th, BYPASS_CAPA, 1);
 425         if (rc != 0)
 426                 GOTO(unlock, rc);
 427
 428         /* 2b. increase child nlink */
 429         rc = dt_ref_add(env, child, th);
 430         dt_write_unlock(env, child);
 431         if (rc != 0)
 432                 GOTO(stop, rc);
 433
 434         /* 3b. insert name into parent dir */
 435         rc = dt_insert(env, parent, (const struct dt_rec *)cfid,
 436                        (const struct dt_key *)name, th, BYPASS_CAPA, 1);
 437         if (rc != 0)
 438                 GOTO(stop, rc);
 439
 440         dt_write_lock(env, parent, 0);
 441         /* 4b. increase parent nlink */
 442         rc = dt_ref_add(env, parent, th);
 443         dt_write_unlock(env, parent);
 444         if (rc != 0)
 445                 GOTO(stop, rc);
 446
 447         bk->lb_lpf_fid = *cfid;
 448         lfsck_bookmark_cpu_to_le(&lfsck->li_bookmark_disk, bk);
 449
 450         /* 5b. update bookmark */
 451         rc = dt_record_write(env, bk_obj,
 452                              lfsck_buf_get(env, bk, len), &pos, th);
 453
 454         GOTO(stop, rc);
 455
 456 unlock:
 457         dt_write_unlock(env, child);
 458
 459 stop:
 460         dt_trans_stop(env, dev, th);
 461
 462         return rc;
 463 }
 464
 465 static int lfsck_create_lpf_remote(const struct lu_env *env,
 466                                    struct lfsck_instance *lfsck,
 467                                    struct dt_object *parent,
 468                                    struct dt_object *child,
 469                                    struct lu_attr *la,
 470                                    struct dt_object_format *dof,
 471                                    const char *name)
 472 {
 473         struct lfsck_bookmark   *bk     = &lfsck->li_bookmark_ram;
 474         struct dt_object        *bk_obj = lfsck->li_bookmark_obj;
 475         const struct lu_fid     *cfid   = lu_object_fid(&child->do_lu);
 476         struct thandle          *th     = NULL;
 477         struct dt_device        *dev;
 478         loff_t                   pos    = 0;
 479         int                      len    = sizeof(struct lfsck_bookmark);
 480         int                      rc     = 0;
 481         ENTRY;
 482
 483         /* Create .lustre/lost+found/MDTxxxx. */
 484
 485         /* XXX: Currently, cross-MDT create operation needs to create the child
 486          *      object firstly, then insert name into the parent directory. For
 487          *      this case, the child object resides on current MDT (local), but
 488          *      the parent ".lustre/lost+found" may be on remote MDT. It is not
 489          *      easy to contain all the sub-modifications orderly within single
 490          *      transaction.
 491          *
 492          *      To avoid more inconsistency, we split the create operation into
 493          *      two transactions:
 494          *
 495          *      1) create the child locally.
 496          *      2) insert the name "MDTXXXX" in the parent ".lustre/lost+found"
 497          *         remotely and update the lfsck_bookmark::lb_lpf_fid locally.
 498          *
 499          *      If 1) done but 2) failed, then the worst case is that we lose
 500          *      one object locally, which is not a big issue. (can be repaird
 501          *      by LFSCK phase III) */
 502
 503         /* Transaction I: */
 504
 505         dev = lfsck->li_bottom;
 506         th = dt_trans_create(env, dev);
 507         if (IS_ERR(th))
 508                 RETURN(PTR_ERR(th));
 509
 510         /* 1a. create child locally. */
 511         rc = dt_declare_create(env, child, la, NULL, dof, th);
 512         if (rc != 0)
 513                 GOTO(stop, rc);
 514
 515         /* 2a. increase child nlink locally. */
 516         rc = dt_declare_ref_add(env, child, th);
 517         if (rc != 0)
 518                 GOTO(stop, rc);
 519
 520         rc = dt_trans_start_local(env, dev, th);
 521         if (rc != 0)
 522                 GOTO(stop, rc);
 523
 524         dt_write_lock(env, child, 0);
 525         /* 1b. create child locally. */
 526         rc = dt_create(env, child, la, NULL, dof, th);
 527         if (rc != 0)
 528                 GOTO(unlock, rc);
 529
 530         if (unlikely(!dt_try_as_dir(env, child)))
 531                 GOTO(unlock, rc = -ENOTDIR);
 532
 533         /* 2b.1 insert dot into child dir locally. */
 534         rc = dt_insert(env, child, (const struct dt_rec *)cfid,
 535                        (const struct dt_key *)dot, th, BYPASS_CAPA, 1);
 536         if (rc != 0)
 537                 GOTO(unlock, rc);
 538
 539         /* 2b.2 insert dotdot into child dir locally. */
 540         rc = dt_insert(env, child, (const struct dt_rec *)&LU_LPF_FID,
 541                        (const struct dt_key *)dotdot, th, BYPASS_CAPA, 1);
 542         if (rc != 0)
 543                 GOTO(unlock, rc);
 544
 545         /* 2b.3 increase child nlink locally. */
 546         rc = dt_ref_add(env, child, th);
 547         dt_write_unlock(env, child);
 548         dt_trans_stop(env, dev, th);
 549         if (rc != 0)
 550                 RETURN(rc);
 551
 552         /* Transaction II: */
 553
 554         dev = lfsck->li_next;
 555         th = dt_trans_create(env, dev);
 556         if (IS_ERR(th))
 557                 RETURN(PTR_ERR(th));
 558
 559         /* 3a. insert name into parent dir remotely. */
 560         rc = dt_declare_insert(env, parent, (const struct dt_rec *)cfid,
 561                                (const struct dt_key *)name, th);
 562         if (rc != 0)
 563                 GOTO(stop, rc);
 564
 565         /* 4a. increase parent nlink remotely. */
 566         rc = dt_declare_ref_add(env, parent, th);
 567         if (rc != 0)
 568                 GOTO(stop, rc);
 569
 570         /* 5a. decrease child nlink for dotdot locally if former remote
 571          *     update failed. */
 572         rc = dt_declare_ref_del(env, child, th);
 573         if (rc != 0)
 574                 GOTO(stop, rc);
 575
 576         /* 6a. decrease child nlink for dot locally if former remote
 577          *     update failed. */
 578         rc = dt_declare_ref_del(env, child, th);
 579         if (rc != 0)
 580                 GOTO(stop, rc);
 581
 582         /* 7a. destroy child locally if former remote update failed. */
 583         rc = dt_declare_destroy(env, child, th);
 584         if (rc != 0)
 585                 GOTO(stop, rc);
 586
 587         /* 8a. update bookmark locally. */
 588         rc = dt_declare_record_write(env, bk_obj,
 589                                      lfsck_buf_get(env, bk, len), 0, th);
 590         if (rc != 0)
 591                 GOTO(stop, rc);
 592
 593         rc = dt_trans_start(env, dev, th);
 594         if (rc != 0)
 595                 GOTO(stop, rc);
 596
 597         /* 3b. insert name into parent dir remotely. */
 598         rc = dt_insert(env, parent, (const struct dt_rec *)cfid,
 599                        (const struct dt_key *)name, th, BYPASS_CAPA, 1);
 600         if (rc == 0) {
 601                 dt_write_lock(env, parent, 0);
 602                 /* 4b. increase parent nlink remotely. */
 603                 rc = dt_ref_add(env, parent, th);
 604                 dt_write_unlock(env, parent);
 605         }
 606         if (rc != 0) {
 607                 /* 5b. decrease child nlink for dotdot locally. */
 608                 dt_ref_del(env, child, th);
 609                 /* 6b. decrease child nlink for dot locally. */
 610                 dt_ref_del(env, child, th);
 611                 /* 7b. destroy child locally. */
 612                 dt_destroy(env, child, th);
 613                 GOTO(stop, rc);
 614         }
 615
 616         bk->lb_lpf_fid = *cfid;
 617         lfsck_bookmark_cpu_to_le(&lfsck->li_bookmark_disk, bk);
 618
 619         /* 8b. update bookmark locally. */
 620         rc = dt_record_write(env, bk_obj,
 621                              lfsck_buf_get(env, bk, len), &pos, th);
 622
 623         GOTO(stop, rc);
 624
 625 unlock:
 626         dt_write_unlock(env, child);
 627 stop:
 628         dt_trans_stop(env, dev, th);
 629
 630         return rc;
 631 }
 632
 633 /* Do NOT create .lustre/lost+found/MDTxxxx when register the lfsck instance,
 634  * because the MDT0 maybe not reaady for sequence allocation yet. We do that
 635  * only when it is required, such as orphan OST-objects repairing. */
 636 int lfsck_create_lpf(const struct lu_env *env, struct lfsck_instance *lfsck)
 637 {
 638         struct lfsck_bookmark    *bk    = &lfsck->li_bookmark_ram;
 639         struct lfsck_thread_info *info  = lfsck_env_info(env);
 640         struct lu_fid            *cfid  = &info->lti_fid2;
 641         struct lu_attr           *la    = &info->lti_la;
 642         struct dt_object_format  *dof   = &info->lti_dof;
 643         struct dt_object         *parent = NULL;
 644         struct dt_object         *child = NULL;
 645         char                      name[8];
 646         int                       node  = lfsck_dev_idx(lfsck->li_bottom);
 647         int                       rc    = 0;
 648         ENTRY;
 649
 650         LASSERT(lfsck->li_master);
 651
 652         sprintf(name, "MDT%04x", node);
 653         if (node == 0) {
 654                 parent = lfsck_object_find_by_dev(env, lfsck->li_bottom,
 655                                                   &LU_LPF_FID);
 656         } else {
 657                 struct lfsck_tgt_desc *ltd;
 658
 659                 ltd = lfsck_tgt_get(&lfsck->li_mdt_descs, 0);
 660                 if (unlikely(ltd == NULL))
 661                         RETURN(-ENODEV);
 662
 663                 parent = lfsck_object_find_by_dev(env, ltd->ltd_tgt,
 664                                                   &LU_LPF_FID);
 665                 lfsck_tgt_put(ltd);
 666         }
 667         if (IS_ERR(parent))
 668                 RETURN(PTR_ERR(parent));
 669
 670         if (unlikely(!dt_try_as_dir(env, parent)))
 671                 GOTO(out, rc = -ENOTDIR);
 672
 673         mutex_lock(&lfsck->li_mutex);
 674         if (lfsck->li_lpf_obj != NULL)
 675                 GOTO(unlock, rc = 0);
 676
 677         if (fid_is_zero(&bk->lb_lpf_fid)) {
 678                 /* There is corner case that: in former LFSCK scanning we have
 679                  * created the .lustre/lost+found/MDTxxxx but failed to update
 680                  * the lfsck_bookmark::lb_lpf_fid successfully. So need lookup
 681                  * it from MDT0 firstly. */
 682                 rc = dt_lookup(env, parent, (struct dt_rec *)cfid,
 683                                (const struct dt_key *)name, BYPASS_CAPA);
 684                 if (rc != 0 && rc != -ENOENT)
 685                         GOTO(unlock, rc);
 686
 687                 if (rc == 0) {
 688                         bk->lb_lpf_fid = *cfid;
 689                         rc = lfsck_bookmark_store(env, lfsck);
 690                 } else {
 691                         rc = lfsck_fid_alloc(env, lfsck, cfid, true);
 692                 }
 693                 if (rc != 0)
 694                         GOTO(unlock, rc);
 695         } else {
 696                 *cfid = bk->lb_lpf_fid;
 697         }
 698
 699         child = lfsck_object_find_by_dev(env, lfsck->li_bottom, cfid);
 700         if (IS_ERR(child))
 701                 GOTO(unlock, rc = PTR_ERR(child));
 702
 703         if (dt_object_exists(child) != 0) {
 704                 if (unlikely(!dt_try_as_dir(env, child)))
 705                         GOTO(unlock, rc = -ENOTDIR);
 706
 707                 lfsck->li_lpf_obj = child;
 708                 GOTO(unlock, rc = 0);
 709         }
 710
 711         memset(la, 0, sizeof(*la));
 712         la->la_atime = la->la_mtime = la->la_ctime = cfs_time_current_sec();
 713         la->la_mode = S_IFDIR | S_IRWXU;
 714         la->la_valid = LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE |
 715                        LA_UID | LA_GID;
 716         memset(dof, 0, sizeof(*dof));
 717         dof->dof_type = dt_mode_to_dft(S_IFDIR);
 718
 719         if (node == 0)
 720                 rc = lfsck_create_lpf_local(env, lfsck, parent, child, la,
 721                                             dof, name);
 722         else
 723                 rc = lfsck_create_lpf_remote(env, lfsck, parent, child, la,
 724                                              dof, name);
 725         if (rc == 0)
 726                 lfsck->li_lpf_obj = child;
 727
 728         GOTO(unlock, rc);
 729
 730 unlock:
 731         mutex_unlock(&lfsck->li_mutex);
 732         if (rc != 0 && child != NULL && !IS_ERR(child))
 733                 lu_object_put(env, &child->do_lu);
 734 out:
 735         if (parent != NULL && !IS_ERR(parent))
 736                 lu_object_put(env, &parent->do_lu);
 737
 738         return rc;
 739 }
 740
 741 static int lfsck_fid_init(struct lfsck_instance *lfsck)
 742 {
 743         struct lfsck_bookmark   *bk     = &lfsck->li_bookmark_ram;
 744         struct seq_server_site  *ss;
 745         char                    *prefix;
 746         int                      rc     = 0;
 747         ENTRY;
 748
 749         ss = lu_site2seq(lfsck->li_bottom->dd_lu_dev.ld_site);
 750         if (unlikely(ss == NULL))
 751                 RETURN(-ENODEV);
 752
 753         OBD_ALLOC_PTR(lfsck->li_seq);
 754         if (lfsck->li_seq == NULL)
 755                 RETURN(-ENOMEM);
 756
 757         OBD_ALLOC(prefix, MAX_OBD_NAME + 7);
 758         if (prefix == NULL)
 759                 GOTO(out, rc = -ENOMEM);
 760
 761         snprintf(prefix, MAX_OBD_NAME + 7, "lfsck-%s", lfsck_lfsck2name(lfsck));
 762         rc = seq_client_init(lfsck->li_seq, NULL, LUSTRE_SEQ_METADATA, prefix,
 763                              ss->ss_server_seq);
 764         OBD_FREE(prefix, MAX_OBD_NAME + 7);
 765         if (rc != 0)
 766                 GOTO(out, rc);
 767
 768         if (fid_is_sane(&bk->lb_last_fid))
 769                 lfsck->li_seq->lcs_fid = bk->lb_last_fid;
 770
 771         RETURN(0);
 772
 773 out:
 774         OBD_FREE_PTR(lfsck->li_seq);
 775         lfsck->li_seq = NULL;
 776
 777         return rc;
 778 }
 779
 780 static void lfsck_fid_fini(struct lfsck_instance *lfsck)
 781 {
 782         if (lfsck->li_seq != NULL) {
 783                 seq_client_fini(lfsck->li_seq);
 784                 OBD_FREE_PTR(lfsck->li_seq);
 785                 lfsck->li_seq = NULL;
 786         }
 787 }
 788
 789 void lfsck_instance_cleanup(const struct lu_env *env,
 790                             struct lfsck_instance *lfsck)
 791 {
 792         struct ptlrpc_thread    *thread = &lfsck->li_thread;
 793         struct lfsck_component  *com;
 794         ENTRY;
 795
 796         LASSERT(list_empty(&lfsck->li_link));
 797         LASSERT(thread_is_init(thread) || thread_is_stopped(thread));
 798
 799         if (lfsck->li_obj_oit != NULL) {
 800                 lu_object_put_nocache(env, &lfsck->li_obj_oit->do_lu);
 801                 lfsck->li_obj_oit = NULL;
 802         }
 803
 804         LASSERT(lfsck->li_obj_dir == NULL);
 805
 806         while (!cfs_list_empty(&lfsck->li_list_scan)) {
 807                 com = cfs_list_entry(lfsck->li_list_scan.next,
 808                                      struct lfsck_component,
 809                                      lc_link);
 810                 lfsck_component_cleanup(env, com);
 811         }
 812
 813         LASSERT(cfs_list_empty(&lfsck->li_list_dir));
 814
 815         while (!cfs_list_empty(&lfsck->li_list_double_scan)) {
 816                 com = cfs_list_entry(lfsck->li_list_double_scan.next,
 817                                      struct lfsck_component,
 818                                      lc_link);
 819                 lfsck_component_cleanup(env, com);
 820         }
 821
 822         while (!cfs_list_empty(&lfsck->li_list_idle)) {
 823                 com = cfs_list_entry(lfsck->li_list_idle.next,
 824                                      struct lfsck_component,
 825                                      lc_link);
 826                 lfsck_component_cleanup(env, com);
 827         }
 828
 829         lfsck_tgt_descs_fini(&lfsck->li_ost_descs);
 830         lfsck_tgt_descs_fini(&lfsck->li_mdt_descs);
 831
 832         if (lfsck->li_bookmark_obj != NULL) {
 833                 lu_object_put_nocache(env, &lfsck->li_bookmark_obj->do_lu);
 834                 lfsck->li_bookmark_obj = NULL;
 835         }
 836
 837         if (lfsck->li_lpf_obj != NULL) {
 838                 lu_object_put(env, &lfsck->li_lpf_obj->do_lu);
 839                 lfsck->li_lpf_obj = NULL;
 840         }
 841
 842         if (lfsck->li_los != NULL) {
 843                 local_oid_storage_fini(env, lfsck->li_los);
 844                 lfsck->li_los = NULL;
 845         }
 846
 847         lfsck_fid_fini(lfsck);
 848
 849         OBD_FREE_PTR(lfsck);
 850 }
 851
 852 static inline struct lfsck_instance *
 853 __lfsck_instance_find(struct dt_device *key, bool ref, bool unlink)
 854 {
 855         struct lfsck_instance *lfsck;
 856
 857         cfs_list_for_each_entry(lfsck, &lfsck_instance_list, li_link) {
 858                 if (lfsck->li_bottom == key) {
 859                         if (ref)
 860                                 lfsck_instance_get(lfsck);
 861                         if (unlink)
 862                                 list_del_init(&lfsck->li_link);
 863
 864                         return lfsck;
 865                 }
 866         }
 867
 868         return NULL;
 869 }
 870
 871 struct lfsck_instance *lfsck_instance_find(struct dt_device *key, bool ref,
 872                                            bool unlink)
 873 {
 874         struct lfsck_instance *lfsck;
 875
 876         spin_lock(&lfsck_instance_lock);
 877         lfsck = __lfsck_instance_find(key, ref, unlink);
 878         spin_unlock(&lfsck_instance_lock);
 879
 880         return lfsck;
 881 }
 882
 883 static inline int lfsck_instance_add(struct lfsck_instance *lfsck)
 884 {
 885         struct lfsck_instance *tmp;
 886
 887         spin_lock(&lfsck_instance_lock);
 888         cfs_list_for_each_entry(tmp, &lfsck_instance_list, li_link) {
 889                 if (lfsck->li_bottom == tmp->li_bottom) {
 890                         spin_unlock(&lfsck_instance_lock);
 891                         return -EEXIST;
 892                 }
 893         }
 894
 895         cfs_list_add_tail(&lfsck->li_link, &lfsck_instance_list);
 896         spin_unlock(&lfsck_instance_lock);
 897         return 0;
 898 }
 899
 900 int lfsck_bits_dump(char **buf, int *len, int bits, const char *names[],
 901                     const char *prefix)
 902 {
 903         int save = *len;
 904         int flag;
 905         int rc;
 906         int i;
 907
 908         rc = snprintf(*buf, *len, "%s:%c", prefix, bits != 0 ? ' ' : '\n');
 909         if (rc <= 0)
 910                 return -ENOSPC;
 911
 912         *buf += rc;
 913         *len -= rc;
 914         for (i = 0, flag = 1; bits != 0; i++, flag = 1 << i) {
 915                 if (flag & bits) {
 916                         bits &= ~flag;
 917                         if (names[i] != NULL) {
 918                                 rc = snprintf(*buf, *len, "%s%c", names[i],
 919                                               bits != 0 ? ',' : '\n');
 920                                 if (rc <= 0)
 921                                         return -ENOSPC;
 922
 923                                 *buf += rc;
 924                                 *len -= rc;
 925                         }
 926                 }
 927         }
 928         return save - *len;
 929 }
 930
 931 int lfsck_time_dump(char **buf, int *len, __u64 time, const char *prefix)
 932 {
 933         int rc;
 934
 935         if (time != 0)
 936                 rc = snprintf(*buf, *len, "%s: "LPU64" seconds\n", prefix,
 937                               cfs_time_current_sec() - time);
 938         else
 939                 rc = snprintf(*buf, *len, "%s: N/A\n", prefix);
 940         if (rc <= 0)
 941                 return -ENOSPC;
 942
 943         *buf += rc;
 944         *len -= rc;
 945         return rc;
 946 }
 947
 948 int lfsck_pos_dump(char **buf, int *len, struct lfsck_position *pos,
 949                    const char *prefix)
 950 {
 951         int rc;
 952
 953         if (fid_is_zero(&pos->lp_dir_parent)) {
 954                 if (pos->lp_oit_cookie == 0)
 955                         rc = snprintf(*buf, *len, "%s: N/A, N/A, N/A\n",
 956                                       prefix);
 957                 else
 958                         rc = snprintf(*buf, *len, "%s: "LPU64", N/A, N/A\n",
 959                                       prefix, pos->lp_oit_cookie);
 960         } else {
 961                 rc = snprintf(*buf, *len, "%s: "LPU64", "DFID", "LPU64"\n",
 962                               prefix, pos->lp_oit_cookie,
 963                               PFID(&pos->lp_dir_parent), pos->lp_dir_cookie);
 964         }
 965         if (rc <= 0)
 966                 return -ENOSPC;
 967
 968         *buf += rc;
 969         *len -= rc;
 970         return rc;
 971 }
 972
 973 void lfsck_pos_fill(const struct lu_env *env, struct lfsck_instance *lfsck,
 974                     struct lfsck_position *pos, bool init)
 975 {
 976         const struct dt_it_ops *iops = &lfsck->li_obj_oit->do_index_ops->dio_it;
 977
 978         if (unlikely(lfsck->li_di_oit == NULL)) {
 979                 memset(pos, 0, sizeof(*pos));
 980                 return;
 981         }
 982
 983         pos->lp_oit_cookie = iops->store(env, lfsck->li_di_oit);
 984         if (!lfsck->li_current_oit_processed && !init)
 985                 pos->lp_oit_cookie--;
 986
 987         LASSERT(pos->lp_oit_cookie > 0);
 988
 989         if (lfsck->li_di_dir != NULL) {
 990                 struct dt_object *dto = lfsck->li_obj_dir;
 991
 992                 pos->lp_dir_cookie = dto->do_index_ops->dio_it.store(env,
 993                                                         lfsck->li_di_dir);
 994
 995                 if (pos->lp_dir_cookie >= MDS_DIR_END_OFF) {
 996                         fid_zero(&pos->lp_dir_parent);
 997                         pos->lp_dir_cookie = 0;
 998                 } else {
 999                         pos->lp_dir_parent = *lfsck_dto2fid(dto);
1000                 }
1001         } else {
1002                 fid_zero(&pos->lp_dir_parent);
1003                 pos->lp_dir_cookie = 0;
1004         }
1005 }
1006
1007 static void __lfsck_set_speed(struct lfsck_instance *lfsck, __u32 limit)
1008 {
1009         lfsck->li_bookmark_ram.lb_speed_limit = limit;
1010         if (limit != LFSCK_SPEED_NO_LIMIT) {
1011                 if (limit > HZ) {
1012                         lfsck->li_sleep_rate = limit / HZ;
1013                         lfsck->li_sleep_jif = 1;
1014                 } else {
1015                         lfsck->li_sleep_rate = 1;
1016                         lfsck->li_sleep_jif = HZ / limit;
1017                 }
1018         } else {
1019                 lfsck->li_sleep_jif = 0;
1020                 lfsck->li_sleep_rate = 0;
1021         }
1022 }
1023
1024 void lfsck_control_speed(struct lfsck_instance *lfsck)
1025 {
1026         struct ptlrpc_thread *thread = &lfsck->li_thread;
1027         struct l_wait_info    lwi;
1028
1029         if (lfsck->li_sleep_jif > 0 &&
1030             lfsck->li_new_scanned >= lfsck->li_sleep_rate) {
1031                 lwi = LWI_TIMEOUT_INTR(lfsck->li_sleep_jif, NULL,
1032                                        LWI_ON_SIGNAL_NOOP, NULL);
1033
1034                 l_wait_event(thread->t_ctl_waitq,
1035                              !thread_is_running(thread),
1036                              &lwi);
1037                 lfsck->li_new_scanned = 0;
1038         }
1039 }
1040
1041 void lfsck_control_speed_by_self(struct lfsck_component *com)
1042 {
1043         struct lfsck_instance   *lfsck  = com->lc_lfsck;
1044         struct ptlrpc_thread    *thread = &lfsck->li_thread;
1045         struct l_wait_info       lwi;
1046
1047         if (lfsck->li_sleep_jif > 0 &&
1048             com->lc_new_scanned >= lfsck->li_sleep_rate) {
1049                 lwi = LWI_TIMEOUT_INTR(lfsck->li_sleep_jif, NULL,
1050                                        LWI_ON_SIGNAL_NOOP, NULL);
1051
1052                 l_wait_event(thread->t_ctl_waitq,
1053                              !thread_is_running(thread),
1054                              &lwi);
1055                 com->lc_new_scanned = 0;
1056         }
1057 }
1058
1059 static int lfsck_parent_fid(const struct lu_env *env, struct dt_object *obj,
1060                             struct lu_fid *fid)
1061 {
1062         if (unlikely(!S_ISDIR(lfsck_object_type(obj)) ||
1063                      !dt_try_as_dir(env, obj)))
1064                 return -ENOTDIR;
1065
1066         return dt_lookup(env, obj, (struct dt_rec *)fid,
1067                          (const struct dt_key *)"..", BYPASS_CAPA);
1068 }
1069
1070 static int lfsck_needs_scan_dir(const struct lu_env *env,
1071                                 struct lfsck_instance *lfsck,
1072                                 struct dt_object *obj)
1073 {
1074         struct lu_fid *fid   = &lfsck_env_info(env)->lti_fid;
1075         int            depth = 0;
1076         int            rc;
1077
1078         if (!lfsck->li_master || !S_ISDIR(lfsck_object_type(obj)) ||
1079             cfs_list_empty(&lfsck->li_list_dir))
1080                RETURN(0);
1081
1082         while (1) {
1083                 /* XXX: Currently, we do not scan the "/REMOTE_PARENT_DIR",
1084                  *      which is the agent directory to manage the objects
1085                  *      which name entries reside on remote MDTs. Related
1086                  *      consistency verification will be processed in LFSCK
1087                  *      phase III. */
1088                 if (lu_fid_eq(lfsck_dto2fid(obj), &lfsck->li_global_root_fid)) {
1089                         if (depth > 0)
1090                                 lfsck_object_put(env, obj);
1091                         return 1;
1092                 }
1093
1094                 /* .lustre doesn't contain "real" user objects, no need lfsck */
1095                 if (fid_is_dot_lustre(lfsck_dto2fid(obj))) {
1096                         if (depth > 0)
1097                                 lfsck_object_put(env, obj);
1098                         return 0;
1099                 }
1100
1101                 dt_read_lock(env, obj, MOR_TGT_CHILD);
1102                 if (unlikely(lfsck_is_dead_obj(obj))) {
1103                         dt_read_unlock(env, obj);
1104                         if (depth > 0)
1105                                 lfsck_object_put(env, obj);
1106                         return 0;
1107                 }
1108
1109                 rc = dt_xattr_get(env, obj,
1110                                   lfsck_buf_get(env, NULL, 0), XATTR_NAME_LINK,
1111                                   BYPASS_CAPA);
1112                 dt_read_unlock(env, obj);
1113                 if (rc >= 0) {
1114                         if (depth > 0)
1115                                 lfsck_object_put(env, obj);
1116                         return 1;
1117                 }
1118
1119                 if (rc < 0 && rc != -ENODATA) {
1120                         if (depth > 0)
1121                                 lfsck_object_put(env, obj);
1122                         return rc;
1123                 }
1124
1125                 rc = lfsck_parent_fid(env, obj, fid);
1126                 if (depth > 0)
1127                         lfsck_object_put(env, obj);
1128                 if (rc != 0)
1129                         return rc;
1130
1131                 if (unlikely(lu_fid_eq(fid, &lfsck->li_local_root_fid)))
1132                         return 0;
1133
1134                 obj = lfsck_object_find(env, lfsck, fid);
1135                 if (obj == NULL)
1136                         return 0;
1137                 else if (IS_ERR(obj))
1138                         return PTR_ERR(obj);
1139
1140                 if (!dt_object_exists(obj)) {
1141                         lfsck_object_put(env, obj);
1142                         return 0;
1143                 }
1144
1145                 /* Currently, only client visible directory can be remote. */
1146                 if (dt_object_remote(obj)) {
1147                         lfsck_object_put(env, obj);
1148                         return 1;
1149                 }
1150
1151                 depth++;
1152         }
1153         return 0;
1154 }
1155
1156 struct lfsck_thread_args *lfsck_thread_args_init(struct lfsck_instance *lfsck,
1157                                                  struct lfsck_component *com,
1158                                                  struct lfsck_start_param *lsp)
1159 {
1160         struct lfsck_thread_args *lta;
1161         int                       rc;
1162
1163         OBD_ALLOC_PTR(lta);
1164         if (lta == NULL)
1165                 return ERR_PTR(-ENOMEM);
1166
1167         rc = lu_env_init(&lta->lta_env, LCT_MD_THREAD | LCT_DT_THREAD);
1168         if (rc != 0) {
1169                 OBD_FREE_PTR(lta);
1170                 return ERR_PTR(rc);
1171         }
1172
1173         lta->lta_lfsck = lfsck_instance_get(lfsck);
1174         if (com != NULL)
1175                 lta->lta_com = lfsck_component_get(com);
1176
1177         lta->lta_lsp = lsp;
1178
1179         return lta;
1180 }
1181
1182 void lfsck_thread_args_fini(struct lfsck_thread_args *lta)
1183 {
1184         if (lta->lta_com != NULL)
1185                 lfsck_component_put(&lta->lta_env, lta->lta_com);
1186         lfsck_instance_put(&lta->lta_env, lta->lta_lfsck);
1187         lu_env_fini(&lta->lta_env);
1188         OBD_FREE_PTR(lta);
1189 }
1190
1191 /* LFSCK wrap functions */
1192
1193 void lfsck_fail(const struct lu_env *env, struct lfsck_instance *lfsck,
1194                 bool new_checked)
1195 {
1196         struct lfsck_component *com;
1197
1198         cfs_list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
1199                 com->lc_ops->lfsck_fail(env, com, new_checked);
1200         }
1201 }
1202
1203 int lfsck_checkpoint(const struct lu_env *env, struct lfsck_instance *lfsck)
1204 {
1205         struct lfsck_component *com;
1206         int                     rc  = 0;
1207         int                     rc1 = 0;
1208
1209         if (likely(cfs_time_beforeq(cfs_time_current(),
1210                                     lfsck->li_time_next_checkpoint)))
1211                 return 0;
1212
1213         lfsck_pos_fill(env, lfsck, &lfsck->li_pos_current, false);
1214         cfs_list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
1215                 rc = com->lc_ops->lfsck_checkpoint(env, com, false);
1216                 if (rc != 0)
1217                         rc1 = rc;
1218         }
1219
1220         lfsck->li_time_last_checkpoint = cfs_time_current();
1221         lfsck->li_time_next_checkpoint = lfsck->li_time_last_checkpoint +
1222                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
1223         return rc1 != 0 ? rc1 : rc;
1224 }
1225
1226 int lfsck_prep(const struct lu_env *env, struct lfsck_instance *lfsck,
1227                struct lfsck_start_param *lsp)
1228 {
1229         struct dt_object       *obj     = NULL;
1230         struct lfsck_component *com;
1231         struct lfsck_component *next;
1232         struct lfsck_position  *pos     = NULL;
1233         const struct dt_it_ops *iops    =
1234                                 &lfsck->li_obj_oit->do_index_ops->dio_it;
1235         struct dt_it           *di;
1236         int                     rc;
1237         ENTRY;
1238
1239         LASSERT(lfsck->li_obj_dir == NULL);
1240         LASSERT(lfsck->li_di_dir == NULL);
1241
1242         lfsck->li_current_oit_processed = 0;
1243         cfs_list_for_each_entry_safe(com, next, &lfsck->li_list_scan, lc_link) {
1244                 com->lc_new_checked = 0;
1245                 if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1246                         com->lc_journal = 0;
1247
1248                 rc = com->lc_ops->lfsck_prep(env, com, lsp);
1249                 if (rc != 0)
1250                         GOTO(out, rc);
1251
1252                 if ((pos == NULL) ||
1253                     (!lfsck_pos_is_zero(&com->lc_pos_start) &&
1254                      lfsck_pos_is_eq(pos, &com->lc_pos_start) > 0))
1255                         pos = &com->lc_pos_start;
1256         }
1257
1258         /* Init otable-based iterator. */
1259         if (pos == NULL) {
1260                 rc = iops->load(env, lfsck->li_di_oit, 0);
1261                 if (rc > 0) {
1262                         lfsck->li_oit_over = 1;
1263                         rc = 0;
1264                 }
1265
1266                 GOTO(out, rc);
1267         }
1268
1269         rc = iops->load(env, lfsck->li_di_oit, pos->lp_oit_cookie);
1270         if (rc < 0)
1271                 GOTO(out, rc);
1272         else if (rc > 0)
1273                 lfsck->li_oit_over = 1;
1274
1275         if (!lfsck->li_master || fid_is_zero(&pos->lp_dir_parent))
1276                 GOTO(out, rc = 0);
1277
1278         /* Find the directory for namespace-based traverse. */
1279         obj = lfsck_object_find(env, lfsck, &pos->lp_dir_parent);
1280         if (obj == NULL)
1281                 GOTO(out, rc = 0);
1282         else if (IS_ERR(obj))
1283                 RETURN(PTR_ERR(obj));
1284
1285         /* XXX: Currently, skip remote object, the consistency for
1286          *      remote object will be processed in LFSCK phase III. */
1287         if (!dt_object_exists(obj) || dt_object_remote(obj) ||
1288             unlikely(!S_ISDIR(lfsck_object_type(obj))))
1289                 GOTO(out, rc = 0);
1290
1291         if (unlikely(!dt_try_as_dir(env, obj)))
1292                 GOTO(out, rc = -ENOTDIR);
1293
1294         /* Init the namespace-based directory traverse. */
1295         iops = &obj->do_index_ops->dio_it;
1296         di = iops->init(env, obj, lfsck->li_args_dir, BYPASS_CAPA);
1297         if (IS_ERR(di))
1298                 GOTO(out, rc = PTR_ERR(di));
1299
1300         LASSERT(pos->lp_dir_cookie < MDS_DIR_END_OFF);
1301
1302         rc = iops->load(env, di, pos->lp_dir_cookie);
1303         if ((rc == 0) || (rc > 0 && pos->lp_dir_cookie > 0))
1304                 rc = iops->next(env, di);
1305         else if (rc > 0)
1306                 rc = 0;
1307
1308         if (rc != 0) {
1309                 iops->put(env, di);
1310                 iops->fini(env, di);
1311                 GOTO(out, rc);
1312         }
1313
1314         lfsck->li_obj_dir = lfsck_object_get(obj);
1315         lfsck->li_cookie_dir = iops->store(env, di);
1316         spin_lock(&lfsck->li_lock);
1317         lfsck->li_di_dir = di;
1318         spin_unlock(&lfsck->li_lock);
1319
1320         GOTO(out, rc = 0);
1321
1322 out:
1323         if (obj != NULL)
1324                 lfsck_object_put(env, obj);
1325
1326         if (rc < 0) {
1327                 cfs_list_for_each_entry_safe(com, next, &lfsck->li_list_scan,
1328                                              lc_link)
1329                         com->lc_ops->lfsck_post(env, com, rc, true);
1330
1331                 return rc;
1332         }
1333
1334         rc = 0;
1335         lfsck_pos_fill(env, lfsck, &lfsck->li_pos_current, true);
1336         cfs_list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
1337                 rc = com->lc_ops->lfsck_checkpoint(env, com, true);
1338                 if (rc != 0)
1339                         break;
1340         }
1341
1342         lfsck->li_time_last_checkpoint = cfs_time_current();
1343         lfsck->li_time_next_checkpoint = lfsck->li_time_last_checkpoint +
1344                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
1345         return rc;
1346 }
1347
1348 int lfsck_exec_oit(const struct lu_env *env, struct lfsck_instance *lfsck,
1349                    struct dt_object *obj)
1350 {
1351         struct lfsck_component *com;
1352         const struct dt_it_ops *iops;
1353         struct dt_it           *di;
1354         int                     rc;
1355         ENTRY;
1356
1357         LASSERT(lfsck->li_obj_dir == NULL);
1358
1359         cfs_list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
1360                 rc = com->lc_ops->lfsck_exec_oit(env, com, obj);
1361                 if (rc != 0)
1362                         RETURN(rc);
1363         }
1364
1365         rc = lfsck_needs_scan_dir(env, lfsck, obj);
1366         if (rc <= 0)
1367                 GOTO(out, rc);
1368
1369         if (unlikely(!dt_try_as_dir(env, obj)))
1370                 GOTO(out, rc = -ENOTDIR);
1371
1372         iops = &obj->do_index_ops->dio_it;
1373         di = iops->init(env, obj, lfsck->li_args_dir, BYPASS_CAPA);
1374         if (IS_ERR(di))
1375                 GOTO(out, rc = PTR_ERR(di));
1376
1377         rc = iops->load(env, di, 0);
1378         if (rc == 0)
1379                 rc = iops->next(env, di);
1380         else if (rc > 0)
1381                 rc = 0;
1382
1383         if (rc != 0) {
1384                 iops->put(env, di);
1385                 iops->fini(env, di);
1386                 GOTO(out, rc);
1387         }
1388
1389         lfsck->li_obj_dir = lfsck_object_get(obj);
1390         lfsck->li_cookie_dir = iops->store(env, di);
1391         spin_lock(&lfsck->li_lock);
1392         lfsck->li_di_dir = di;
1393         spin_unlock(&lfsck->li_lock);
1394
1395         GOTO(out, rc = 0);
1396
1397 out:
1398         if (rc < 0)
1399                 lfsck_fail(env, lfsck, false);
1400         return (rc > 0 ? 0 : rc);
1401 }
1402
1403 int lfsck_exec_dir(const struct lu_env *env, struct lfsck_instance *lfsck,
1404                    struct dt_object *obj, struct lu_dirent *ent)
1405 {
1406         struct lfsck_component *com;
1407         int                     rc;
1408
1409         cfs_list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
1410                 rc = com->lc_ops->lfsck_exec_dir(env, com, obj, ent);
1411                 if (rc != 0)
1412                         return rc;
1413         }
1414         return 0;
1415 }
1416
1417 int lfsck_post(const struct lu_env *env, struct lfsck_instance *lfsck,
1418                int result)
1419 {
1420         struct lfsck_component *com;
1421         struct lfsck_component *next;
1422         int                     rc  = 0;
1423         int                     rc1 = 0;
1424
1425         lfsck_pos_fill(env, lfsck, &lfsck->li_pos_current, false);
1426         cfs_list_for_each_entry_safe(com, next, &lfsck->li_list_scan, lc_link) {
1427                 rc = com->lc_ops->lfsck_post(env, com, result, false);
1428                 if (rc != 0)
1429                         rc1 = rc;
1430         }
1431
1432         lfsck->li_time_last_checkpoint = cfs_time_current();
1433         lfsck->li_time_next_checkpoint = lfsck->li_time_last_checkpoint +
1434                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
1435
1436         /* Ignore some component post failure to make other can go ahead. */
1437         return result;
1438 }
1439
1440 static void lfsck_interpret(const struct lu_env *env,
1441                             struct lfsck_instance *lfsck,
1442                             struct ptlrpc_request *req, void *args, int result)
1443 {
1444         struct lfsck_async_interpret_args *laia = args;
1445         struct lfsck_component            *com;
1446
1447         LASSERT(laia->laia_shared);
1448
1449         spin_lock(&lfsck->li_lock);
1450         list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
1451                 if (com->lc_ops->lfsck_interpret != NULL) {
1452                         laia->laia_com = com;
1453                         com->lc_ops->lfsck_interpret(env, req, laia, result);
1454                 }
1455         }
1456
1457         list_for_each_entry(com, &lfsck->li_list_double_scan, lc_link) {
1458                 if (com->lc_ops->lfsck_interpret != NULL) {
1459                         laia->laia_com = com;
1460                         com->lc_ops->lfsck_interpret(env, req, laia, result);
1461                 }
1462         }
1463         spin_unlock(&lfsck->li_lock);
1464 }
1465
1466 int lfsck_double_scan(const struct lu_env *env, struct lfsck_instance *lfsck)
1467 {
1468         struct lfsck_component *com;
1469         struct lfsck_component *next;
1470         struct l_wait_info      lwi = { 0 };
1471         int                     rc  = 0;
1472         int                     rc1 = 0;
1473
1474         cfs_list_for_each_entry_safe(com, next, &lfsck->li_list_double_scan,
1475                                      lc_link) {
1476                 if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1477                         com->lc_journal = 0;
1478
1479                 rc = com->lc_ops->lfsck_double_scan(env, com);
1480                 if (rc != 0)
1481                         rc1 = rc;
1482         }
1483
1484         l_wait_event(lfsck->li_thread.t_ctl_waitq,
1485                      atomic_read(&lfsck->li_double_scan_count) == 0,
1486                      &lwi);
1487
1488         return rc1 != 0 ? rc1 : rc;
1489 }
1490
1491 static int lfsck_stop_notify(const struct lu_env *env,
1492                              struct lfsck_instance *lfsck,
1493                              struct lfsck_tgt_descs *ltds,
1494                              struct lfsck_tgt_desc *ltd, __u16 type)
1495 {
1496         struct ptlrpc_request_set *set;
1497         struct lfsck_component    *com;
1498         int                        rc  = 0;
1499         ENTRY;
1500
1501         spin_lock(&lfsck->li_lock);
1502         com = __lfsck_component_find(lfsck, type, &lfsck->li_list_scan);
1503         if (com == NULL)
1504                 com = __lfsck_component_find(lfsck, type,
1505                                              &lfsck->li_list_double_scan);
1506         if (com != NULL)
1507                 lfsck_component_get(com);
1508         spin_lock(&lfsck->li_lock);
1509
1510         if (com != NULL) {
1511                 if (com->lc_ops->lfsck_stop_notify != NULL) {
1512                         set = ptlrpc_prep_set();
1513                         if (set == NULL) {
1514                                 lfsck_component_put(env, com);
1515
1516                                 RETURN(-ENOMEM);
1517                         }
1518
1519                         rc = com->lc_ops->lfsck_stop_notify(env, com, ltds,
1520                                                             ltd, set);
1521                         if (rc == 0)
1522                                 rc = ptlrpc_set_wait(set);
1523
1524                         ptlrpc_set_destroy(set);
1525                 }
1526
1527                 lfsck_component_put(env, com);
1528         }
1529
1530         RETURN(rc);
1531 }
1532
1533 void lfsck_quit(const struct lu_env *env, struct lfsck_instance *lfsck)
1534 {
1535         struct lfsck_component *com;
1536         struct lfsck_component *next;
1537
1538         list_for_each_entry_safe(com, next, &lfsck->li_list_scan,
1539                                  lc_link) {
1540                 if (com->lc_ops->lfsck_quit != NULL)
1541                         com->lc_ops->lfsck_quit(env, com);
1542         }
1543
1544         list_for_each_entry_safe(com, next, &lfsck->li_list_double_scan,
1545                                  lc_link) {
1546                 if (com->lc_ops->lfsck_quit != NULL)
1547                         com->lc_ops->lfsck_quit(env, com);
1548         }
1549 }
1550
1551 static int lfsck_async_interpret(const struct lu_env *env,
1552                                  struct ptlrpc_request *req,
1553                                  void *args, int rc)
1554 {
1555         struct lfsck_async_interpret_args *laia = args;
1556         struct lfsck_instance             *lfsck;
1557
1558         lfsck = container_of0(laia->laia_ltds, struct lfsck_instance,
1559                               li_mdt_descs);
1560         lfsck_interpret(env, lfsck, req, laia, rc);
1561         lfsck_tgt_put(laia->laia_ltd);
1562         if (rc != 0 && laia->laia_result != -EALREADY)
1563                 laia->laia_result = rc;
1564
1565         return 0;
1566 }
1567
1568 int lfsck_async_request(const struct lu_env *env, struct obd_export *exp,
1569                         struct lfsck_request *lr,
1570                         struct ptlrpc_request_set *set,
1571                         ptlrpc_interpterer_t interpreter,
1572                         void *args, int request)
1573 {
1574         struct lfsck_async_interpret_args *laia;
1575         struct ptlrpc_request             *req;
1576         struct lfsck_request              *tmp;
1577         struct req_format                 *format;
1578         int                                rc;
1579
1580         if (!(exp_connect_flags(exp) & OBD_CONNECT_LFSCK))
1581                 return -EOPNOTSUPP;
1582
1583         switch (request) {
1584         case LFSCK_NOTIFY:
1585                 format = &RQF_LFSCK_NOTIFY;
1586                 break;
1587         case LFSCK_QUERY:
1588                 format = &RQF_LFSCK_QUERY;
1589                 break;
1590         default:
1591                 CERROR("%s: unknown async request: opc = %d\n",
1592                        exp->exp_obd->obd_name, request);
1593                 return -EINVAL;
1594         }
1595
1596         req = ptlrpc_request_alloc(class_exp2cliimp(exp), format);
1597         if (req == NULL)
1598                 return -ENOMEM;
1599
1600         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, request);
1601         if (rc != 0) {
1602                 ptlrpc_request_free(req);
1603
1604                 return rc;
1605         }
1606
1607         tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
1608         *tmp = *lr;
1609         ptlrpc_request_set_replen(req);
1610
1611         laia = ptlrpc_req_async_args(req);
1612         *laia = *(struct lfsck_async_interpret_args *)args;
1613         if (laia->laia_com != NULL)
1614                 lfsck_component_get(laia->laia_com);
1615         req->rq_interpret_reply = interpreter;
1616         ptlrpc_set_add_req(set, req);
1617
1618         return 0;
1619 }
1620
1621 /* external interfaces */
1622
1623 int lfsck_get_speed(struct dt_device *key, void *buf, int len)
1624 {
1625         struct lu_env           env;
1626         struct lfsck_instance  *lfsck;
1627         int                     rc;
1628         ENTRY;
1629
1630         rc = lu_env_init(&env, LCT_MD_THREAD | LCT_DT_THREAD);
1631         if (rc != 0)
1632                 RETURN(rc);
1633
1634         lfsck = lfsck_instance_find(key, true, false);
1635         if (likely(lfsck != NULL)) {
1636                 rc = snprintf(buf, len, "%u\n",
1637                               lfsck->li_bookmark_ram.lb_speed_limit);
1638                 lfsck_instance_put(&env, lfsck);
1639         } else {
1640                 rc = -ENODEV;
1641         }
1642
1643         lu_env_fini(&env);
1644
1645         RETURN(rc);
1646 }
1647 EXPORT_SYMBOL(lfsck_get_speed);
1648
1649 int lfsck_set_speed(struct dt_device *key, int val)
1650 {
1651         struct lu_env           env;
1652         struct lfsck_instance  *lfsck;
1653         int                     rc;
1654         ENTRY;
1655
1656         rc = lu_env_init(&env, LCT_MD_THREAD | LCT_DT_THREAD);
1657         if (rc != 0)
1658                 RETURN(rc);
1659
1660         lfsck = lfsck_instance_find(key, true, false);
1661         if (likely(lfsck != NULL)) {
1662                 mutex_lock(&lfsck->li_mutex);
1663                 __lfsck_set_speed(lfsck, val);
1664                 rc = lfsck_bookmark_store(&env, lfsck);
1665                 mutex_unlock(&lfsck->li_mutex);
1666                 lfsck_instance_put(&env, lfsck);
1667         } else {
1668                 rc = -ENODEV;
1669         }
1670
1671         lu_env_fini(&env);
1672
1673         RETURN(rc);
1674 }
1675 EXPORT_SYMBOL(lfsck_set_speed);
1676
1677 int lfsck_get_windows(struct dt_device *key, void *buf, int len)
1678 {
1679         struct lu_env           env;
1680         struct lfsck_instance  *lfsck;
1681         int                     rc;
1682         ENTRY;
1683
1684         rc = lu_env_init(&env, LCT_MD_THREAD | LCT_DT_THREAD);
1685         if (rc != 0)
1686                 RETURN(rc);
1687
1688         lfsck = lfsck_instance_find(key, true, false);
1689         if (likely(lfsck != NULL)) {
1690                 rc = snprintf(buf, len, "%u\n",
1691                               lfsck->li_bookmark_ram.lb_async_windows);
1692                 lfsck_instance_put(&env, lfsck);
1693         } else {
1694                 rc = -ENODEV;
1695         }
1696
1697         lu_env_fini(&env);
1698
1699         RETURN(rc);
1700 }
1701 EXPORT_SYMBOL(lfsck_get_windows);
1702
1703 int lfsck_set_windows(struct dt_device *key, int val)
1704 {
1705         struct lu_env           env;
1706         struct lfsck_instance  *lfsck;
1707         int                     rc;
1708         ENTRY;
1709
1710         rc = lu_env_init(&env, LCT_MD_THREAD | LCT_DT_THREAD);
1711         if (rc != 0)
1712                 RETURN(rc);
1713
1714         lfsck = lfsck_instance_find(key, true, false);
1715         if (likely(lfsck != NULL)) {
1716                 if (val > LFSCK_ASYNC_WIN_MAX) {
1717                         CERROR("%s: Too large async windows size, which "
1718                                "may cause memory issues. The valid range "
1719                                "is [0 - %u]. If you do not want to restrict "
1720                                "the windows size for async requests pipeline, "
1721                                "just set it as 0.\n",
1722                                lfsck_lfsck2name(lfsck), LFSCK_ASYNC_WIN_MAX);
1723                         rc = -EINVAL;
1724                 } else if (lfsck->li_bookmark_ram.lb_async_windows != val) {
1725                         mutex_lock(&lfsck->li_mutex);
1726                         lfsck->li_bookmark_ram.lb_async_windows = val;
1727                         rc = lfsck_bookmark_store(&env, lfsck);
1728                         mutex_unlock(&lfsck->li_mutex);
1729                 }
1730                 lfsck_instance_put(&env, lfsck);
1731         } else {
1732                 rc = -ENODEV;
1733         }
1734
1735         lu_env_fini(&env);
1736
1737         RETURN(rc);
1738 }
1739 EXPORT_SYMBOL(lfsck_set_windows);
1740
1741 int lfsck_dump(struct dt_device *key, void *buf, int len, enum lfsck_type type)
1742 {
1743         struct lu_env           env;
1744         struct lfsck_instance  *lfsck;
1745         struct lfsck_component *com;
1746         int                     rc;
1747         ENTRY;
1748
1749         rc = lu_env_init(&env, LCT_MD_THREAD | LCT_DT_THREAD);
1750         if (rc != 0)
1751                 RETURN(rc);
1752
1753         lfsck = lfsck_instance_find(key, true, false);
1754         if (likely(lfsck != NULL)) {
1755                 com = lfsck_component_find(lfsck, type);
1756                 if (likely(com != NULL)) {
1757                         rc = com->lc_ops->lfsck_dump(&env, com, buf, len);
1758                         lfsck_component_put(&env, com);
1759                 } else {
1760                         rc = -ENOTSUPP;
1761                 }
1762
1763                 lfsck_instance_put(&env, lfsck);
1764         } else {
1765                 rc = -ENODEV;
1766         }
1767
1768         lu_env_fini(&env);
1769
1770         RETURN(rc);
1771 }
1772 EXPORT_SYMBOL(lfsck_dump);
1773
1774 static int lfsck_stop_all(const struct lu_env *env,
1775                           struct lfsck_instance *lfsck,
1776                           struct lfsck_stop *stop)
1777 {
1778         struct lfsck_thread_info          *info   = lfsck_env_info(env);
1779         struct lfsck_request              *lr     = &info->lti_lr;
1780         struct lfsck_async_interpret_args *laia   = &info->lti_laia;
1781         struct ptlrpc_request_set         *set;
1782         struct lfsck_tgt_descs            *ltds   = &lfsck->li_mdt_descs;
1783         struct lfsck_tgt_desc             *ltd;
1784         struct lfsck_bookmark             *bk     = &lfsck->li_bookmark_ram;
1785         __u32                              idx;
1786         int                                rc     = 0;
1787         int                                rc1    = 0;
1788         ENTRY;
1789
1790         LASSERT(stop->ls_flags & LPF_BROADCAST);
1791
1792         set = ptlrpc_prep_set();
1793         if (unlikely(set == NULL)) {
1794                 CERROR("%s: cannot allocate memory for stop LFSCK on "
1795                        "all targets\n", lfsck_lfsck2name(lfsck));
1796
1797                 RETURN(-ENOMEM);
1798         }
1799
1800         memset(lr, 0, sizeof(*lr));
1801         lr->lr_event = LE_STOP;
1802         lr->lr_index = lfsck_dev_idx(lfsck->li_bottom);
1803         lr->lr_status = stop->ls_status;
1804         lr->lr_version = bk->lb_version;
1805         lr->lr_active = LFSCK_TYPES_ALL;
1806         lr->lr_param = stop->ls_flags;
1807
1808         laia->laia_com = NULL;
1809         laia->laia_ltds = ltds;
1810         laia->laia_lr = lr;
1811         laia->laia_result = 0;
1812         laia->laia_shared = 1;
1813
1814         down_read(&ltds->ltd_rw_sem);
1815         cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
1816                 ltd = lfsck_tgt_get(ltds, idx);
1817                 LASSERT(ltd != NULL);
1818
1819                 laia->laia_ltd = ltd;
1820                 rc = lfsck_async_request(env, ltd->ltd_exp, lr, set,
1821                                          lfsck_async_interpret, laia,
1822                                          LFSCK_NOTIFY);
1823                 if (rc != 0) {
1824                         lfsck_interpret(env, lfsck, NULL, laia, rc);
1825                         lfsck_tgt_put(ltd);
1826                         CWARN("%s: cannot notify MDT %x for LFSCK stop: "
1827                               "rc = %d\n", lfsck_lfsck2name(lfsck), idx, rc);
1828                         rc1 = rc;
1829                 }
1830         }
1831         up_read(&ltds->ltd_rw_sem);
1832
1833         rc = ptlrpc_set_wait(set);
1834         ptlrpc_set_destroy(set);
1835
1836         if (rc == 0)
1837                 rc = laia->laia_result;
1838
1839         if (rc == -EALREADY)
1840                 rc = 0;
1841
1842         if (rc != 0)
1843                 CWARN("%s: fail to stop LFSCK on some MDTs: rc = %d\n",
1844                       lfsck_lfsck2name(lfsck), rc);
1845
1846         RETURN(rc != 0 ? rc : rc1);
1847 }
1848
1849 static int lfsck_start_all(const struct lu_env *env,
1850                            struct lfsck_instance *lfsck,
1851                            struct lfsck_start *start)
1852 {
1853         struct lfsck_thread_info          *info   = lfsck_env_info(env);
1854         struct lfsck_request              *lr     = &info->lti_lr;
1855         struct lfsck_async_interpret_args *laia   = &info->lti_laia;
1856         struct ptlrpc_request_set         *set;
1857         struct lfsck_tgt_descs            *ltds   = &lfsck->li_mdt_descs;
1858         struct lfsck_tgt_desc             *ltd;
1859         struct lfsck_bookmark             *bk     = &lfsck->li_bookmark_ram;
1860         __u32                              idx;
1861         int                                rc     = 0;
1862         ENTRY;
1863
1864         LASSERT(start->ls_flags & LPF_BROADCAST);
1865
1866         set = ptlrpc_prep_set();
1867         if (unlikely(set == NULL)) {
1868                 if (bk->lb_param & LPF_FAILOUT) {
1869                         CERROR("%s: cannot allocate memory for start LFSCK on "
1870                                "all targets, failout.\n",
1871                                lfsck_lfsck2name(lfsck));
1872
1873                         RETURN(-ENOMEM);
1874                 } else {
1875                         CWARN("%s: cannot allocate memory for start LFSCK on "
1876                               "all targets, partly scan.\n",
1877                               lfsck_lfsck2name(lfsck));
1878
1879                         RETURN(0);
1880                 }
1881         }
1882
1883         memset(lr, 0, sizeof(*lr));
1884         lr->lr_event = LE_START;
1885         lr->lr_index = lfsck_dev_idx(lfsck->li_bottom);
1886         lr->lr_speed = bk->lb_speed_limit;
1887         lr->lr_version = bk->lb_version;
1888         lr->lr_active = start->ls_active;
1889         lr->lr_param = start->ls_flags;
1890         lr->lr_async_windows = bk->lb_async_windows;
1891         lr->lr_valid = LSV_SPEED_LIMIT | LSV_ERROR_HANDLE | LSV_DRYRUN |
1892                        LSV_ASYNC_WINDOWS;
1893
1894         laia->laia_com = NULL;
1895         laia->laia_ltds = ltds;
1896         laia->laia_lr = lr;
1897         laia->laia_result = 0;
1898         laia->laia_shared = 1;
1899
1900         down_read(&ltds->ltd_rw_sem);
1901         cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
1902                 ltd = lfsck_tgt_get(ltds, idx);
1903                 LASSERT(ltd != NULL);
1904
1905                 laia->laia_ltd = ltd;
1906                 ltd->ltd_layout_done = 0;
1907                 rc = lfsck_async_request(env, ltd->ltd_exp, lr, set,
1908                                          lfsck_async_interpret, laia,
1909                                          LFSCK_NOTIFY);
1910                 if (rc != 0) {
1911                         lfsck_interpret(env, lfsck, NULL, laia, rc);
1912                         lfsck_tgt_put(ltd);
1913                         if (bk->lb_param & LPF_FAILOUT) {
1914                                 CERROR("%s: cannot notify MDT %x for LFSCK "
1915                                        "start, failout: rc = %d\n",
1916                                        lfsck_lfsck2name(lfsck), idx, rc);
1917                                 break;
1918                         } else {
1919                                 CWARN("%s: cannot notify MDT %x for LFSCK "
1920                                       "start, partly scan: rc = %d\n",
1921                                       lfsck_lfsck2name(lfsck), idx, rc);
1922                                 rc = 0;
1923                         }
1924                 }
1925         }
1926         up_read(&ltds->ltd_rw_sem);
1927
1928         if (rc != 0) {
1929                 ptlrpc_set_destroy(set);
1930
1931                 RETURN(rc);
1932         }
1933
1934         rc = ptlrpc_set_wait(set);
1935         ptlrpc_set_destroy(set);
1936
1937         if (rc == 0)
1938                 rc = laia->laia_result;
1939
1940         if (rc != 0) {
1941                 if (bk->lb_param & LPF_FAILOUT) {
1942                         struct lfsck_stop *stop = &info->lti_stop;
1943
1944                         CERROR("%s: cannot start LFSCK on some MDTs, "
1945                                "stop all: rc = %d\n",
1946                                lfsck_lfsck2name(lfsck), rc);
1947                         if (rc != -EALREADY) {
1948                                 stop->ls_status = LS_FAILED;
1949                                 stop->ls_flags = LPF_ALL_TGT | LPF_BROADCAST;
1950                                 lfsck_stop_all(env, lfsck, stop);
1951                         }
1952                 } else {
1953                         CWARN("%s: cannot start LFSCK on some MDTs, "
1954                               "partly scan: rc = %d\n",
1955                               lfsck_lfsck2name(lfsck), rc);
1956                         rc = 0;
1957                 }
1958         }
1959
1960         RETURN(rc);
1961 }
1962
1963 int lfsck_start(const struct lu_env *env, struct dt_device *key,
1964                 struct lfsck_start_param *lsp)
1965 {
1966         struct lfsck_start              *start  = lsp->lsp_start;
1967         struct lfsck_instance           *lfsck;
1968         struct lfsck_bookmark           *bk;
1969         struct ptlrpc_thread            *thread;
1970         struct lfsck_component          *com;
1971         struct l_wait_info               lwi    = { 0 };
1972         struct lfsck_thread_args        *lta;
1973         bool                             dirty  = false;
1974         long                             rc     = 0;
1975         __u16                            valid  = 0;
1976         __u16                            flags  = 0;
1977         __u16                            type   = 1;
1978         ENTRY;
1979
1980         lfsck = lfsck_instance_find(key, true, false);
1981         if (unlikely(lfsck == NULL))
1982                 RETURN(-ENODEV);
1983
1984         /* System is not ready, try again later. */
1985         if (unlikely(lfsck->li_namespace == NULL))
1986                 GOTO(put, rc = -EAGAIN);
1987
1988         /* start == NULL means auto trigger paused LFSCK. */
1989         if ((start == NULL) &&
1990             (cfs_list_empty(&lfsck->li_list_scan) ||
1991              OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NO_AUTO)))
1992                 GOTO(put, rc = 0);
1993
1994         bk = &lfsck->li_bookmark_ram;
1995         thread = &lfsck->li_thread;
1996         mutex_lock(&lfsck->li_mutex);
1997         spin_lock(&lfsck->li_lock);
1998         if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
1999                 rc = -EALREADY;
2000                 while (start->ls_active != 0) {
2001                         if (!(type & start->ls_active)) {
2002                                 type <<= 1;
2003                                 continue;
2004                         }
2005
2006                         com = __lfsck_component_find(lfsck, type,
2007                                                      &lfsck->li_list_scan);
2008                         if (com == NULL)
2009                                 com = __lfsck_component_find(lfsck, type,
2010                                                 &lfsck->li_list_double_scan);
2011                         if (com == NULL) {
2012                                 rc = -EOPNOTSUPP;
2013                                 break;
2014                         }
2015
2016                         if (com->lc_ops->lfsck_join != NULL) {
2017                                 rc = com->lc_ops->lfsck_join( env, com, lsp);
2018                                 if (rc != 0 && rc != -EALREADY)
2019                                         break;
2020                         }
2021                         start->ls_active &= ~type;
2022                         type <<= 1;
2023                 }
2024                 spin_unlock(&lfsck->li_lock);
2025                 GOTO(out, rc);
2026         }
2027         spin_unlock(&lfsck->li_lock);
2028
2029         lfsck->li_status = 0;
2030         lfsck->li_oit_over = 0;
2031         lfsck->li_start_unplug = 0;
2032         lfsck->li_drop_dryrun = 0;
2033         lfsck->li_new_scanned = 0;
2034
2035         /* For auto trigger. */
2036         if (start == NULL)
2037                 goto trigger;
2038
2039         if (start->ls_flags & LPF_BROADCAST && !lfsck->li_master) {
2040                 CERROR("%s: only allow to specify '-A | -o' via MDS\n",
2041                        lfsck_lfsck2name(lfsck));
2042
2043                 GOTO(out, rc = -EPERM);
2044         }
2045
2046         start->ls_version = bk->lb_version;
2047         if (start->ls_valid & LSV_SPEED_LIMIT) {
2048                 __lfsck_set_speed(lfsck, start->ls_speed_limit);
2049                 dirty = true;
2050         }
2051
2052         if (start->ls_valid & LSV_ASYNC_WINDOWS &&
2053             bk->lb_async_windows != start->ls_async_windows) {
2054                 bk->lb_async_windows = start->ls_async_windows;
2055                 dirty = true;
2056         }
2057
2058         if (start->ls_valid & LSV_ERROR_HANDLE) {
2059                 valid |= DOIV_ERROR_HANDLE;
2060                 if (start->ls_flags & LPF_FAILOUT)
2061                         flags |= DOIF_FAILOUT;
2062
2063                 if ((start->ls_flags & LPF_FAILOUT) &&
2064                     !(bk->lb_param & LPF_FAILOUT)) {
2065                         bk->lb_param |= LPF_FAILOUT;
2066                         dirty = true;
2067                 } else if (!(start->ls_flags & LPF_FAILOUT) &&
2068                            (bk->lb_param & LPF_FAILOUT)) {
2069                         bk->lb_param &= ~LPF_FAILOUT;
2070                         dirty = true;
2071                 }
2072         }
2073
2074         if (start->ls_valid & LSV_DRYRUN) {
2075                 valid |= DOIV_DRYRUN;
2076                 if (start->ls_flags & LPF_DRYRUN)
2077                         flags |= DOIF_DRYRUN;
2078
2079                 if ((start->ls_flags & LPF_DRYRUN) &&
2080                     !(bk->lb_param & LPF_DRYRUN)) {
2081                         bk->lb_param |= LPF_DRYRUN;
2082                         dirty = true;
2083                 } else if (!(start->ls_flags & LPF_DRYRUN) &&
2084                            (bk->lb_param & LPF_DRYRUN)) {
2085                         bk->lb_param &= ~LPF_DRYRUN;
2086                         lfsck->li_drop_dryrun = 1;
2087                         dirty = true;
2088                 }
2089         }
2090
2091         if (bk->lb_param & LPF_ALL_TGT &&
2092             !(start->ls_flags & LPF_ALL_TGT)) {
2093                 bk->lb_param &= ~LPF_ALL_TGT;
2094                 dirty = true;
2095         } else if (!(bk->lb_param & LPF_ALL_TGT) &&
2096                    start->ls_flags & LPF_ALL_TGT) {
2097                 bk->lb_param |= LPF_ALL_TGT;
2098                 dirty = true;
2099         }
2100
2101         if (bk->lb_param & LPF_ORPHAN &&
2102             !(start->ls_flags & LPF_ORPHAN)) {
2103                 bk->lb_param &= ~LPF_ORPHAN;
2104                 dirty = true;
2105         } else if (!(bk->lb_param & LPF_ORPHAN) &&
2106                    start->ls_flags & LPF_ORPHAN) {
2107                 bk->lb_param |= LPF_ORPHAN;
2108                 dirty = true;
2109         }
2110
2111         if (dirty) {
2112                 rc = lfsck_bookmark_store(env, lfsck);
2113                 if (rc != 0)
2114                         GOTO(out, rc);
2115         }
2116
2117         if (start->ls_flags & LPF_RESET)
2118                 flags |= DOIF_RESET;
2119
2120         if (start->ls_active != 0) {
2121                 struct lfsck_component *next;
2122
2123                 if (start->ls_active == LFSCK_TYPES_ALL)
2124                         start->ls_active = LFSCK_TYPES_SUPPORTED;
2125
2126                 if (start->ls_active & ~LFSCK_TYPES_SUPPORTED) {
2127                         start->ls_active &= ~LFSCK_TYPES_SUPPORTED;
2128                         GOTO(out, rc = -ENOTSUPP);
2129                 }
2130
2131                 cfs_list_for_each_entry_safe(com, next,
2132                                              &lfsck->li_list_scan, lc_link) {
2133                         if (!(com->lc_type & start->ls_active)) {
2134                                 rc = com->lc_ops->lfsck_post(env, com, 0,
2135                                                              false);
2136                                 if (rc != 0)
2137                                         GOTO(out, rc);
2138                         }
2139                 }
2140
2141                 while (start->ls_active != 0) {
2142                         if (type & start->ls_active) {
2143                                 com = __lfsck_component_find(lfsck, type,
2144                                                         &lfsck->li_list_idle);
2145                                 if (com != NULL) {
2146                                         /* The component status will be updated
2147                                          * when its prep() is called later by
2148                                          * the LFSCK main engine. */
2149                                         cfs_list_del_init(&com->lc_link);
2150                                         cfs_list_add_tail(&com->lc_link,
2151                                                           &lfsck->li_list_scan);
2152                                 }
2153                                 start->ls_active &= ~type;
2154                         }
2155                         type <<= 1;
2156                 }
2157         }
2158
2159         cfs_list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
2160                 start->ls_active |= com->lc_type;
2161                 if (flags & DOIF_RESET) {
2162                         rc = com->lc_ops->lfsck_reset(env, com, false);
2163                         if (rc != 0)
2164                                 GOTO(out, rc);
2165                 }
2166         }
2167
2168 trigger:
2169         lfsck->li_args_dir = LUDA_64BITHASH | LUDA_VERIFY;
2170         if (bk->lb_param & LPF_DRYRUN) {
2171                 lfsck->li_args_dir |= LUDA_VERIFY_DRYRUN;
2172                 valid |= DOIV_DRYRUN;
2173                 flags |= DOIF_DRYRUN;
2174         }
2175
2176         if (bk->lb_param & LPF_FAILOUT) {
2177                 valid |= DOIV_ERROR_HANDLE;
2178                 flags |= DOIF_FAILOUT;
2179         }
2180
2181         if (!cfs_list_empty(&lfsck->li_list_scan))
2182                 flags |= DOIF_OUTUSED;
2183
2184         lfsck->li_args_oit = (flags << DT_OTABLE_IT_FLAGS_SHIFT) | valid;
2185         thread_set_flags(thread, 0);
2186         lta = lfsck_thread_args_init(lfsck, NULL, lsp);
2187         if (IS_ERR(lta))
2188                 GOTO(out, rc = PTR_ERR(lta));
2189
2190         rc = PTR_ERR(kthread_run(lfsck_master_engine, lta, "lfsck"));
2191         if (IS_ERR_VALUE(rc)) {
2192                 CERROR("%s: cannot start LFSCK thread: rc = %ld\n",
2193                        lfsck_lfsck2name(lfsck), rc);
2194                 lfsck_thread_args_fini(lta);
2195
2196                 GOTO(out, rc);
2197         }
2198
2199         l_wait_event(thread->t_ctl_waitq,
2200                      thread_is_running(thread) ||
2201                      thread_is_stopped(thread),
2202                      &lwi);
2203         if (start == NULL || !(start->ls_flags & LPF_BROADCAST)) {
2204                 lfsck->li_start_unplug = 1;
2205                 wake_up_all(&thread->t_ctl_waitq);
2206
2207                 GOTO(out, rc = 0);
2208         }
2209
2210         /* release lfsck::li_mutex to avoid deadlock. */
2211         mutex_unlock(&lfsck->li_mutex);
2212         rc = lfsck_start_all(env, lfsck, start);
2213         if (rc != 0) {
2214                 spin_lock(&lfsck->li_lock);
2215                 if (thread_is_stopped(thread)) {
2216                         spin_unlock(&lfsck->li_lock);
2217                 } else {
2218                         lfsck->li_status = LS_FAILED;
2219                         lfsck->li_flags = 0;
2220                         thread_set_flags(thread, SVC_STOPPING);
2221                         spin_unlock(&lfsck->li_lock);
2222
2223                         lfsck->li_start_unplug = 1;
2224                         wake_up_all(&thread->t_ctl_waitq);
2225                         l_wait_event(thread->t_ctl_waitq,
2226                                      thread_is_stopped(thread),
2227                                      &lwi);
2228                 }
2229         } else {
2230                 lfsck->li_start_unplug = 1;
2231                 wake_up_all(&thread->t_ctl_waitq);
2232         }
2233
2234         GOTO(put, rc);
2235
2236 out:
2237         mutex_unlock(&lfsck->li_mutex);
2238
2239 put:
2240         lfsck_instance_put(env, lfsck);
2241
2242         return rc < 0 ? rc : 0;
2243 }
2244 EXPORT_SYMBOL(lfsck_start);
2245
2246 int lfsck_stop(const struct lu_env *env, struct dt_device *key,
2247                struct lfsck_stop *stop)
2248 {
2249         struct lfsck_instance   *lfsck;
2250         struct ptlrpc_thread    *thread;
2251         struct l_wait_info       lwi    = { 0 };
2252         int                      rc     = 0;
2253         int                      rc1    = 0;
2254         ENTRY;
2255
2256         lfsck = lfsck_instance_find(key, true, false);
2257         if (unlikely(lfsck == NULL))
2258                 RETURN(-ENODEV);
2259
2260         thread = &lfsck->li_thread;
2261         /* release lfsck::li_mutex to avoid deadlock. */
2262         if (stop != NULL && stop->ls_flags & LPF_BROADCAST) {
2263                 if (!lfsck->li_master) {
2264                         CERROR("%s: only allow to specify '-A' via MDS\n",
2265                                lfsck_lfsck2name(lfsck));
2266
2267                         GOTO(out, rc = -EPERM);
2268                 }
2269
2270                 rc1 = lfsck_stop_all(env, lfsck, stop);
2271         }
2272
2273         mutex_lock(&lfsck->li_mutex);
2274         spin_lock(&lfsck->li_lock);
2275         if (thread_is_init(thread) || thread_is_stopped(thread)) {
2276                 spin_unlock(&lfsck->li_lock);
2277                 GOTO(out, rc = -EALREADY);
2278         }
2279
2280         if (stop != NULL) {
2281                 lfsck->li_status = stop->ls_status;
2282                 lfsck->li_flags = stop->ls_flags;
2283         } else {
2284                 lfsck->li_status = LS_STOPPED;
2285                 lfsck->li_flags = 0;
2286         }
2287
2288         thread_set_flags(thread, SVC_STOPPING);
2289         spin_unlock(&lfsck->li_lock);
2290
2291         wake_up_all(&thread->t_ctl_waitq);
2292         l_wait_event(thread->t_ctl_waitq,
2293                      thread_is_stopped(thread),
2294                      &lwi);
2295
2296         GOTO(out, rc = 0);
2297
2298 out:
2299         mutex_unlock(&lfsck->li_mutex);
2300         lfsck_instance_put(env, lfsck);
2301
2302         return rc != 0 ? rc : rc1;
2303 }
2304 EXPORT_SYMBOL(lfsck_stop);
2305
2306 int lfsck_in_notify(const struct lu_env *env, struct dt_device *key,
2307                     struct lfsck_request *lr)
2308 {
2309         int rc = -EOPNOTSUPP;
2310         ENTRY;
2311
2312         switch (lr->lr_event) {
2313         case LE_START: {
2314                 struct lfsck_start       *start = &lfsck_env_info(env)->lti_start;
2315                 struct lfsck_start_param  lsp;
2316
2317                 memset(start, 0, sizeof(*start));
2318                 start->ls_valid = lr->lr_valid;
2319                 start->ls_speed_limit = lr->lr_speed;
2320                 start->ls_version = lr->lr_version;
2321                 start->ls_active = lr->lr_active;
2322                 start->ls_flags = lr->lr_param & ~LPF_BROADCAST;
2323                 start->ls_async_windows = lr->lr_async_windows;
2324
2325                 lsp.lsp_start = start;
2326                 lsp.lsp_index = lr->lr_index;
2327                 lsp.lsp_index_valid = 1;
2328                 rc = lfsck_start(env, key, &lsp);
2329                 break;
2330         }
2331         case LE_STOP: {
2332                 struct lfsck_stop *stop = &lfsck_env_info(env)->lti_stop;
2333
2334                 memset(stop, 0, sizeof(*stop));
2335                 stop->ls_status = lr->lr_status;
2336                 stop->ls_flags = lr->lr_param & ~LPF_BROADCAST;
2337                 rc = lfsck_stop(env, key, stop);
2338                 break;
2339         }
2340         case LE_PHASE1_DONE:
2341         case LE_PHASE2_DONE:
2342         case LE_FID_ACCESSED:
2343         case LE_PEER_EXIT:
2344         case LE_CONDITIONAL_DESTROY:
2345         case LE_PAIRS_VERIFY: {
2346                 struct lfsck_instance  *lfsck;
2347                 struct lfsck_component *com;
2348
2349                 lfsck = lfsck_instance_find(key, true, false);
2350                 if (unlikely(lfsck == NULL))
2351                         RETURN(-ENODEV);
2352
2353                 com = lfsck_component_find(lfsck, lr->lr_active);
2354                 if (likely(com != NULL)) {
2355                         rc = com->lc_ops->lfsck_in_notify(env, com, lr);
2356                         lfsck_component_put(env, com);
2357                 }
2358
2359                 lfsck_instance_put(env, lfsck);
2360                 break;
2361         }
2362         default:
2363                 break;
2364         }
2365
2366         RETURN(rc);
2367 }
2368 EXPORT_SYMBOL(lfsck_in_notify);
2369
2370 int lfsck_query(const struct lu_env *env, struct dt_device *key,
2371                 struct lfsck_request *lr)
2372 {
2373         struct lfsck_instance  *lfsck;
2374         struct lfsck_component *com;
2375         int                     rc;
2376         ENTRY;
2377
2378         lfsck = lfsck_instance_find(key, true, false);
2379         if (unlikely(lfsck == NULL))
2380                 RETURN(-ENODEV);
2381
2382         com = lfsck_component_find(lfsck, lr->lr_active);
2383         if (likely(com != NULL)) {
2384                 rc = com->lc_ops->lfsck_query(env, com);
2385                 lfsck_component_put(env, com);
2386         } else {
2387                 rc = -ENOTSUPP;
2388         }
2389
2390         lfsck_instance_put(env, lfsck);
2391
2392         RETURN(rc);
2393 }
2394 EXPORT_SYMBOL(lfsck_query);
2395
2396 int lfsck_register_namespace(const struct lu_env *env, struct dt_device *key,
2397                              struct ldlm_namespace *ns)
2398 {
2399         struct lfsck_instance  *lfsck;
2400         int                     rc      = -ENODEV;
2401
2402         lfsck = lfsck_instance_find(key, true, false);
2403         if (likely(lfsck != NULL)) {
2404                 lfsck->li_namespace = ns;
2405                 lfsck_instance_put(env, lfsck);
2406                 rc = 0;
2407         }
2408
2409         return rc;
2410 }
2411 EXPORT_SYMBOL(lfsck_register_namespace);
2412
2413 int lfsck_register(const struct lu_env *env, struct dt_device *key,
2414                    struct dt_device *next, struct obd_device *obd,
2415                    lfsck_out_notify notify, void *notify_data, bool master)
2416 {
2417         struct lfsck_instance   *lfsck;
2418         struct dt_object        *root  = NULL;
2419         struct dt_object        *obj;
2420         struct lu_fid           *fid   = &lfsck_env_info(env)->lti_fid;
2421         int                      rc;
2422         ENTRY;
2423
2424         lfsck = lfsck_instance_find(key, false, false);
2425         if (unlikely(lfsck != NULL))
2426                 RETURN(-EEXIST);
2427
2428         OBD_ALLOC_PTR(lfsck);
2429         if (lfsck == NULL)
2430                 RETURN(-ENOMEM);
2431
2432         mutex_init(&lfsck->li_mutex);
2433         spin_lock_init(&lfsck->li_lock);
2434         CFS_INIT_LIST_HEAD(&lfsck->li_link);
2435         CFS_INIT_LIST_HEAD(&lfsck->li_list_scan);
2436         CFS_INIT_LIST_HEAD(&lfsck->li_list_dir);
2437         CFS_INIT_LIST_HEAD(&lfsck->li_list_double_scan);
2438         CFS_INIT_LIST_HEAD(&lfsck->li_list_idle);
2439         atomic_set(&lfsck->li_ref, 1);
2440         atomic_set(&lfsck->li_double_scan_count, 0);
2441         init_waitqueue_head(&lfsck->li_thread.t_ctl_waitq);
2442         lfsck->li_out_notify = notify;
2443         lfsck->li_out_notify_data = notify_data;
2444         lfsck->li_next = next;
2445         lfsck->li_bottom = key;
2446         lfsck->li_obd = obd;
2447
2448         rc = lfsck_tgt_descs_init(&lfsck->li_ost_descs);
2449         if (rc != 0)
2450                 GOTO(out, rc);
2451
2452         rc = lfsck_tgt_descs_init(&lfsck->li_mdt_descs);
2453         if (rc != 0)
2454                 GOTO(out, rc);
2455
2456         fid->f_seq = FID_SEQ_LOCAL_NAME;
2457         fid->f_oid = 1;
2458         fid->f_ver = 0;
2459         rc = local_oid_storage_init(env, lfsck->li_bottom, fid, &lfsck->li_los);
2460         if (rc != 0)
2461                 GOTO(out, rc);
2462
2463         rc = dt_root_get(env, key, fid);
2464         if (rc != 0)
2465                 GOTO(out, rc);
2466
2467         root = dt_locate(env, lfsck->li_bottom, fid);
2468         if (IS_ERR(root))
2469                 GOTO(out, rc = PTR_ERR(root));
2470
2471         if (unlikely(!dt_try_as_dir(env, root)))
2472                 GOTO(out, rc = -ENOTDIR);
2473
2474         lfsck->li_local_root_fid = *fid;
2475         if (master) {
2476                 lfsck->li_master = 1;
2477                 if (lfsck_dev_idx(lfsck->li_bottom) == 0) {
2478                         rc = dt_lookup(env, root,
2479                                 (struct dt_rec *)(&lfsck->li_global_root_fid),
2480                                 (const struct dt_key *)"ROOT", BYPASS_CAPA);
2481                         if (rc != 0)
2482                                 GOTO(out, rc);
2483                 }
2484         }
2485
2486         fid->f_seq = FID_SEQ_LOCAL_FILE;
2487         fid->f_oid = OTABLE_IT_OID;
2488         fid->f_ver = 0;
2489         obj = dt_locate(env, lfsck->li_bottom, fid);
2490         if (IS_ERR(obj))
2491                 GOTO(out, rc = PTR_ERR(obj));
2492
2493         lfsck->li_obj_oit = obj;
2494         rc = obj->do_ops->do_index_try(env, obj, &dt_otable_features);
2495         if (rc != 0) {
2496                 if (rc == -ENOTSUPP)
2497                         GOTO(add, rc = 0);
2498
2499                 GOTO(out, rc);
2500         }
2501
2502         rc = lfsck_bookmark_setup(env, lfsck);
2503         if (rc != 0)
2504                 GOTO(out, rc);
2505
2506         if (master) {
2507                 rc = lfsck_fid_init(lfsck);
2508                 if (rc < 0)
2509                         GOTO(out, rc);
2510
2511                 rc = lfsck_namespace_setup(env, lfsck);
2512                 if (rc < 0)
2513                         GOTO(out, rc);
2514         }
2515
2516         rc = lfsck_layout_setup(env, lfsck);
2517         if (rc < 0)
2518                 GOTO(out, rc);
2519
2520         /* XXX: more LFSCK components initialization to be added here. */
2521
2522 add:
2523         rc = lfsck_instance_add(lfsck);
2524         if (rc == 0)
2525                 rc = lfsck_add_target_from_orphan(env, lfsck);
2526 out:
2527         if (root != NULL && !IS_ERR(root))
2528                 lu_object_put(env, &root->do_lu);
2529         if (rc != 0)
2530                 lfsck_instance_cleanup(env, lfsck);
2531         return rc;
2532 }
2533 EXPORT_SYMBOL(lfsck_register);
2534
2535 void lfsck_degister(const struct lu_env *env, struct dt_device *key)
2536 {
2537         struct lfsck_instance *lfsck;
2538
2539         lfsck = lfsck_instance_find(key, false, true);
2540         if (lfsck != NULL)
2541                 lfsck_instance_put(env, lfsck);
2542 }
2543 EXPORT_SYMBOL(lfsck_degister);
2544
2545 int lfsck_add_target(const struct lu_env *env, struct dt_device *key,
2546                      struct dt_device *tgt, struct obd_export *exp,
2547                      __u32 index, bool for_ost)
2548 {
2549         struct lfsck_instance   *lfsck;
2550         struct lfsck_tgt_desc   *ltd;
2551         int                      rc;
2552         ENTRY;
2553
2554         OBD_ALLOC_PTR(ltd);
2555         if (ltd == NULL)
2556                 RETURN(-ENOMEM);
2557
2558         ltd->ltd_tgt = tgt;
2559         ltd->ltd_key = key;
2560         ltd->ltd_exp = exp;
2561         INIT_LIST_HEAD(&ltd->ltd_orphan_list);
2562         INIT_LIST_HEAD(&ltd->ltd_layout_list);
2563         INIT_LIST_HEAD(&ltd->ltd_layout_phase_list);
2564         atomic_set(&ltd->ltd_ref, 1);
2565         ltd->ltd_index = index;
2566
2567         spin_lock(&lfsck_instance_lock);
2568         lfsck = __lfsck_instance_find(key, true, false);
2569         if (lfsck == NULL) {
2570                 if (for_ost)
2571                         list_add_tail(&ltd->ltd_orphan_list,
2572                                       &lfsck_ost_orphan_list);
2573                 else
2574                         list_add_tail(&ltd->ltd_orphan_list,
2575                                       &lfsck_mdt_orphan_list);
2576                 spin_unlock(&lfsck_instance_lock);
2577
2578                 RETURN(0);
2579         }
2580         spin_unlock(&lfsck_instance_lock);
2581
2582         rc = __lfsck_add_target(env, lfsck, ltd, for_ost, false);
2583         if (rc != 0)
2584                 lfsck_tgt_put(ltd);
2585
2586         lfsck_instance_put(env, lfsck);
2587
2588         RETURN(rc);
2589 }
2590 EXPORT_SYMBOL(lfsck_add_target);
2591
2592 void lfsck_del_target(const struct lu_env *env, struct dt_device *key,
2593                       struct dt_device *tgt, __u32 index, bool for_ost)
2594 {
2595         struct lfsck_instance   *lfsck;
2596         struct lfsck_tgt_descs  *ltds;
2597         struct lfsck_tgt_desc   *ltd    = NULL;
2598         struct list_head        *head;
2599
2600         if (for_ost)
2601                 head = &lfsck_ost_orphan_list;
2602         else
2603                 head = &lfsck_mdt_orphan_list;
2604
2605         spin_lock(&lfsck_instance_lock);
2606         list_for_each_entry(ltd, head, ltd_orphan_list) {
2607                 if (ltd->ltd_tgt == tgt) {
2608                         list_del_init(&ltd->ltd_orphan_list);
2609                         spin_unlock(&lfsck_instance_lock);
2610                         lfsck_tgt_put(ltd);
2611
2612                         return;
2613                 }
2614         }
2615
2616         lfsck = __lfsck_instance_find(key, true, false);
2617         spin_unlock(&lfsck_instance_lock);
2618         if (unlikely(lfsck == NULL))
2619                 return;
2620
2621         if (for_ost)
2622                 ltds = &lfsck->li_ost_descs;
2623         else
2624                 ltds = &lfsck->li_mdt_descs;
2625
2626         down_write(&ltds->ltd_rw_sem);
2627         LASSERT(ltds->ltd_tgts_bitmap != NULL);
2628
2629         if (unlikely(index >= ltds->ltd_tgts_bitmap->size))
2630                 goto unlock;
2631
2632         ltd = LTD_TGT(ltds, index);
2633         if (unlikely(ltd == NULL))
2634                 goto unlock;
2635
2636         LASSERT(ltds->ltd_tgtnr > 0);
2637
2638         ltds->ltd_tgtnr--;
2639         cfs_bitmap_clear(ltds->ltd_tgts_bitmap, index);
2640         LTD_TGT(ltds, index) = NULL;
2641
2642 unlock:
2643         if (ltd == NULL) {
2644                 if (for_ost)
2645                         head = &lfsck->li_ost_descs.ltd_orphan;
2646                 else
2647                         head = &lfsck->li_ost_descs.ltd_orphan;
2648
2649                 list_for_each_entry(ltd, head, ltd_orphan_list) {
2650                         if (ltd->ltd_tgt == tgt) {
2651                                 list_del_init(&ltd->ltd_orphan_list);
2652                                 break;
2653                         }
2654                 }
2655         }
2656
2657         up_write(&ltds->ltd_rw_sem);
2658         if (ltd != NULL) {
2659                 spin_lock(&ltds->ltd_lock);
2660                 ltd->ltd_dead = 1;
2661                 spin_unlock(&ltds->ltd_lock);
2662                 lfsck_stop_notify(env, lfsck, ltds, ltd, LT_LAYOUT);
2663                 lfsck_tgt_put(ltd);
2664         }
2665
2666         lfsck_instance_put(env, lfsck);
2667 }
2668 EXPORT_SYMBOL(lfsck_del_target);
2669
2670 static int __init lfsck_init(void)
2671 {
2672         int rc;
2673
2674         INIT_LIST_HEAD(&lfsck_ost_orphan_list);
2675         INIT_LIST_HEAD(&lfsck_mdt_orphan_list);
2676         lfsck_key_init_generic(&lfsck_thread_key, NULL);
2677         rc = lu_context_key_register(&lfsck_thread_key);
2678         if (rc == 0) {
2679                 tgt_register_lfsck_in_notify(lfsck_in_notify);
2680                 tgt_register_lfsck_query(lfsck_query);
2681         }
2682
2683         return rc;
2684 }
2685
2686 static void __exit lfsck_exit(void)
2687 {
2688         struct lfsck_tgt_desc *ltd;
2689         struct lfsck_tgt_desc *next;
2690
2691         LASSERT(cfs_list_empty(&lfsck_instance_list));
2692
2693         list_for_each_entry_safe(ltd, next, &lfsck_ost_orphan_list,
2694                                  ltd_orphan_list) {
2695                 list_del_init(&ltd->ltd_orphan_list);
2696                 lfsck_tgt_put(ltd);
2697         }
2698
2699         list_for_each_entry_safe(ltd, next, &lfsck_mdt_orphan_list,
2700                                  ltd_orphan_list) {
2701                 list_del_init(&ltd->ltd_orphan_list);
2702                 lfsck_tgt_put(ltd);
2703         }
2704
2705         lu_context_key_degister(&lfsck_thread_key);
2706 }
2707
2708 MODULE_AUTHOR("Intel Corporation <http://www.intel.com/>");
2709 MODULE_DESCRIPTION("LFSCK");
2710 MODULE_LICENSE("GPL");
2711
2712 cfs_module(lfsck, LUSTRE_VERSION_STRING, lfsck_init, lfsck_exit);