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