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         OBD_FREE_PTR(info);
  55 }
  56
  57 LU_CONTEXT_KEY_DEFINE(lfsck, LCT_MD_THREAD | LCT_DT_THREAD);
  58 LU_KEY_INIT_GENERIC(lfsck);
  59
  60 static CFS_LIST_HEAD(lfsck_instance_list);
  61 static DEFINE_SPINLOCK(lfsck_instance_lock);
  62
  63 const char *lfsck_status_names[] = {
  64         "init",
  65         "scanning-phase1",
  66         "scanning-phase2",
  67         "completed",
  68         "failed",
  69         "stopped",
  70         "paused",
  71         "crashed",
  72         NULL
  73 };
  74
  75 const char *lfsck_flags_names[] = {
  76         "scanned-once",
  77         "inconsistent",
  78         "upgrade",
  79         NULL
  80 };
  81
  82 const char *lfsck_param_names[] = {
  83         "failout",
  84         "dryrun",
  85         NULL
  86 };
  87
  88 static inline mdsno_t lfsck_dev_idx(struct dt_device *dev)
  89 {
  90         return dev->dd_lu_dev.ld_site->ld_seq_site->ss_node_id;
  91 }
  92
  93 static inline void lfsck_component_get(struct lfsck_component *com)
  94 {
  95         atomic_inc(&com->lc_ref);
  96 }
  97
  98 static inline void lfsck_component_put(const struct lu_env *env,
  99                                        struct lfsck_component *com)
 100 {
 101         if (atomic_dec_and_test(&com->lc_ref)) {
 102                 if (com->lc_obj != NULL)
 103                         lu_object_put_nocache(env, &com->lc_obj->do_lu);
 104                 if (com->lc_file_ram != NULL)
 105                         OBD_FREE(com->lc_file_ram, com->lc_file_size);
 106                 if (com->lc_file_disk != NULL)
 107                         OBD_FREE(com->lc_file_disk, com->lc_file_size);
 108                 OBD_FREE_PTR(com);
 109         }
 110 }
 111
 112 static inline struct lfsck_component *
 113 __lfsck_component_find(struct lfsck_instance *lfsck, __u16 type, cfs_list_t *list)
 114 {
 115         struct lfsck_component *com;
 116
 117         cfs_list_for_each_entry(com, list, lc_link) {
 118                 if (com->lc_type == type)
 119                         return com;
 120         }
 121         return NULL;
 122 }
 123
 124 static struct lfsck_component *
 125 lfsck_component_find(struct lfsck_instance *lfsck, __u16 type)
 126 {
 127         struct lfsck_component *com;
 128
 129         spin_lock(&lfsck->li_lock);
 130         com = __lfsck_component_find(lfsck, type, &lfsck->li_list_scan);
 131         if (com != NULL)
 132                 goto unlock;
 133
 134         com = __lfsck_component_find(lfsck, type,
 135                                      &lfsck->li_list_double_scan);
 136         if (com != NULL)
 137                 goto unlock;
 138
 139         com = __lfsck_component_find(lfsck, type, &lfsck->li_list_idle);
 140
 141 unlock:
 142         if (com != NULL)
 143                 lfsck_component_get(com);
 144         spin_unlock(&lfsck->li_lock);
 145         return com;
 146 }
 147
 148 void lfsck_component_cleanup(const struct lu_env *env,
 149                              struct lfsck_component *com)
 150 {
 151         if (!cfs_list_empty(&com->lc_link))
 152                 cfs_list_del_init(&com->lc_link);
 153         if (!cfs_list_empty(&com->lc_link_dir))
 154                 cfs_list_del_init(&com->lc_link_dir);
 155
 156         lfsck_component_put(env, com);
 157 }
 158
 159 static void lfsck_instance_cleanup(const struct lu_env *env,
 160                                    struct lfsck_instance *lfsck)
 161 {
 162         struct ptlrpc_thread    *thread = &lfsck->li_thread;
 163         struct lfsck_component  *com;
 164         ENTRY;
 165
 166         LASSERT(list_empty(&lfsck->li_link));
 167         LASSERT(thread_is_init(thread) || thread_is_stopped(thread));
 168
 169         if (lfsck->li_obj_oit != NULL) {
 170                 lu_object_put_nocache(env, &lfsck->li_obj_oit->do_lu);
 171                 lfsck->li_obj_oit = NULL;
 172         }
 173
 174         LASSERT(lfsck->li_obj_dir == NULL);
 175
 176         while (!cfs_list_empty(&lfsck->li_list_scan)) {
 177                 com = cfs_list_entry(lfsck->li_list_scan.next,
 178                                      struct lfsck_component,
 179                                      lc_link);
 180                 lfsck_component_cleanup(env, com);
 181         }
 182
 183         LASSERT(cfs_list_empty(&lfsck->li_list_dir));
 184
 185         while (!cfs_list_empty(&lfsck->li_list_double_scan)) {
 186                 com = cfs_list_entry(lfsck->li_list_double_scan.next,
 187                                      struct lfsck_component,
 188                                      lc_link);
 189                 lfsck_component_cleanup(env, com);
 190         }
 191
 192         while (!cfs_list_empty(&lfsck->li_list_idle)) {
 193                 com = cfs_list_entry(lfsck->li_list_idle.next,
 194                                      struct lfsck_component,
 195                                      lc_link);
 196                 lfsck_component_cleanup(env, com);
 197         }
 198
 199         if (lfsck->li_bookmark_obj != NULL) {
 200                 lu_object_put_nocache(env, &lfsck->li_bookmark_obj->do_lu);
 201                 lfsck->li_bookmark_obj = NULL;
 202         }
 203
 204         if (lfsck->li_los != NULL) {
 205                 local_oid_storage_fini(env, lfsck->li_los);
 206                 lfsck->li_los = NULL;
 207         }
 208
 209         OBD_FREE_PTR(lfsck);
 210 }
 211
 212 static inline void lfsck_instance_get(struct lfsck_instance *lfsck)
 213 {
 214         atomic_inc(&lfsck->li_ref);
 215 }
 216
 217 static inline void lfsck_instance_put(const struct lu_env *env,
 218                                       struct lfsck_instance *lfsck)
 219 {
 220         if (atomic_dec_and_test(&lfsck->li_ref))
 221                 lfsck_instance_cleanup(env, lfsck);
 222 }
 223
 224 static inline struct lfsck_instance *lfsck_instance_find(struct dt_device *key,
 225                                                          bool ref, bool unlink)
 226 {
 227         struct lfsck_instance *lfsck;
 228
 229         spin_lock(&lfsck_instance_lock);
 230         cfs_list_for_each_entry(lfsck, &lfsck_instance_list, li_link) {
 231                 if (lfsck->li_bottom == key) {
 232                         if (ref)
 233                                 lfsck_instance_get(lfsck);
 234                         if (unlink)
 235                                 list_del_init(&lfsck->li_link);
 236                         spin_unlock(&lfsck_instance_lock);
 237                         return lfsck;
 238                 }
 239         }
 240         spin_unlock(&lfsck_instance_lock);
 241         return NULL;
 242 }
 243
 244 static inline int lfsck_instance_add(struct lfsck_instance *lfsck)
 245 {
 246         struct lfsck_instance *tmp;
 247
 248         spin_lock(&lfsck_instance_lock);
 249         cfs_list_for_each_entry(tmp, &lfsck_instance_list, li_link) {
 250                 if (lfsck->li_bottom == tmp->li_bottom) {
 251                         spin_unlock(&lfsck_instance_lock);
 252                         return -EEXIST;
 253                 }
 254         }
 255
 256         cfs_list_add_tail(&lfsck->li_link, &lfsck_instance_list);
 257         spin_unlock(&lfsck_instance_lock);
 258         return 0;
 259 }
 260
 261 int lfsck_bits_dump(char **buf, int *len, int bits, const char *names[],
 262                     const char *prefix)
 263 {
 264         int save = *len;
 265         int flag;
 266         int rc;
 267         int i;
 268
 269         rc = snprintf(*buf, *len, "%s:%c", prefix, bits != 0 ? ' ' : '\n');
 270         if (rc <= 0)
 271                 return -ENOSPC;
 272
 273         *buf += rc;
 274         *len -= rc;
 275         for (i = 0, flag = 1; bits != 0; i++, flag = 1 << i) {
 276                 if (flag & bits) {
 277                         bits &= ~flag;
 278                         rc = snprintf(*buf, *len, "%s%c", names[i],
 279                                       bits != 0 ? ',' : '\n');
 280                         if (rc <= 0)
 281                                 return -ENOSPC;
 282
 283                         *buf += rc;
 284                         *len -= rc;
 285                 }
 286         }
 287         return save - *len;
 288 }
 289
 290 int lfsck_time_dump(char **buf, int *len, __u64 time, const char *prefix)
 291 {
 292         int rc;
 293
 294         if (time != 0)
 295                 rc = snprintf(*buf, *len, "%s: "LPU64" seconds\n", prefix,
 296                               cfs_time_current_sec() - time);
 297         else
 298                 rc = snprintf(*buf, *len, "%s: N/A\n", prefix);
 299         if (rc <= 0)
 300                 return -ENOSPC;
 301
 302         *buf += rc;
 303         *len -= rc;
 304         return rc;
 305 }
 306
 307 int lfsck_pos_dump(char **buf, int *len, struct lfsck_position *pos,
 308                    const char *prefix)
 309 {
 310         int rc;
 311
 312         if (fid_is_zero(&pos->lp_dir_parent)) {
 313                 if (pos->lp_oit_cookie == 0)
 314                         rc = snprintf(*buf, *len, "%s: N/A, N/A, N/A\n",
 315                                       prefix);
 316                 else
 317                         rc = snprintf(*buf, *len, "%s: "LPU64", N/A, N/A\n",
 318                                       prefix, pos->lp_oit_cookie);
 319         } else {
 320                 rc = snprintf(*buf, *len, "%s: "LPU64", "DFID", "LPU64"\n",
 321                               prefix, pos->lp_oit_cookie,
 322                               PFID(&pos->lp_dir_parent), pos->lp_dir_cookie);
 323         }
 324         if (rc <= 0)
 325                 return -ENOSPC;
 326
 327         *buf += rc;
 328         *len -= rc;
 329         return rc;
 330 }
 331
 332 void lfsck_pos_fill(const struct lu_env *env, struct lfsck_instance *lfsck,
 333                     struct lfsck_position *pos, bool init)
 334 {
 335         const struct dt_it_ops *iops = &lfsck->li_obj_oit->do_index_ops->dio_it;
 336
 337         if (unlikely(lfsck->li_di_oit == NULL)) {
 338                 memset(pos, 0, sizeof(*pos));
 339                 return;
 340         }
 341
 342         pos->lp_oit_cookie = iops->store(env, lfsck->li_di_oit);
 343         if (!lfsck->li_current_oit_processed && !init)
 344                 pos->lp_oit_cookie--;
 345
 346         LASSERT(pos->lp_oit_cookie > 0);
 347
 348         if (lfsck->li_di_dir != NULL) {
 349                 struct dt_object *dto = lfsck->li_obj_dir;
 350
 351                 pos->lp_dir_cookie = dto->do_index_ops->dio_it.store(env,
 352                                                         lfsck->li_di_dir);
 353
 354                 if (pos->lp_dir_cookie >= MDS_DIR_END_OFF) {
 355                         fid_zero(&pos->lp_dir_parent);
 356                         pos->lp_dir_cookie = 0;
 357                 } else {
 358                         pos->lp_dir_parent = *lu_object_fid(&dto->do_lu);
 359                 }
 360         } else {
 361                 fid_zero(&pos->lp_dir_parent);
 362                 pos->lp_dir_cookie = 0;
 363         }
 364 }
 365
 366 static void __lfsck_set_speed(struct lfsck_instance *lfsck, __u32 limit)
 367 {
 368         lfsck->li_bookmark_ram.lb_speed_limit = limit;
 369         if (limit != LFSCK_SPEED_NO_LIMIT) {
 370                 if (limit > HZ) {
 371                         lfsck->li_sleep_rate = limit / HZ;
 372                         lfsck->li_sleep_jif = 1;
 373                 } else {
 374                         lfsck->li_sleep_rate = 1;
 375                         lfsck->li_sleep_jif = HZ / limit;
 376                 }
 377         } else {
 378                 lfsck->li_sleep_jif = 0;
 379                 lfsck->li_sleep_rate = 0;
 380         }
 381 }
 382
 383 void lfsck_control_speed(struct lfsck_instance *lfsck)
 384 {
 385         struct ptlrpc_thread *thread = &lfsck->li_thread;
 386         struct l_wait_info    lwi;
 387
 388         if (lfsck->li_sleep_jif > 0 &&
 389             lfsck->li_new_scanned >= lfsck->li_sleep_rate) {
 390                 spin_lock(&lfsck->li_lock);
 391                 if (likely(lfsck->li_sleep_jif > 0 &&
 392                            lfsck->li_new_scanned >= lfsck->li_sleep_rate)) {
 393                         lwi = LWI_TIMEOUT_INTR(lfsck->li_sleep_jif, NULL,
 394                                                LWI_ON_SIGNAL_NOOP, NULL);
 395                         spin_unlock(&lfsck->li_lock);
 396
 397                         l_wait_event(thread->t_ctl_waitq,
 398                                      !thread_is_running(thread),
 399                                      &lwi);
 400                         lfsck->li_new_scanned = 0;
 401                 } else {
 402                         spin_unlock(&lfsck->li_lock);
 403                 }
 404         }
 405 }
 406
 407 static int lfsck_parent_fid(const struct lu_env *env, struct dt_object *obj,
 408                             struct lu_fid *fid)
 409 {
 410         if (unlikely(!S_ISDIR(lfsck_object_type(obj)) ||
 411                      !dt_try_as_dir(env, obj)))
 412                 return -ENOTDIR;
 413
 414         return dt_lookup(env, obj, (struct dt_rec *)fid,
 415                          (const struct dt_key *)"..", BYPASS_CAPA);
 416 }
 417
 418 static int lfsck_needs_scan_dir(const struct lu_env *env,
 419                                 struct lfsck_instance *lfsck,
 420                                 struct dt_object *obj)
 421 {
 422         struct lu_fid *fid   = &lfsck_env_info(env)->lti_fid;
 423         int            depth = 0;
 424         int            rc;
 425
 426         if (!lfsck->li_master || !S_ISDIR(lfsck_object_type(obj)) ||
 427             cfs_list_empty(&lfsck->li_list_dir))
 428                RETURN(0);
 429
 430         while (1) {
 431                 /* XXX: Currently, we do not scan the "/REMOTE_PARENT_DIR",
 432                  *      which is the agent directory to manage the objects
 433                  *      which name entries reside on remote MDTs. Related
 434                  *      consistency verification will be processed in LFSCK
 435                  *      phase III. */
 436                 if (lu_fid_eq(lfsck_dto2fid(obj), &lfsck->li_global_root_fid)) {
 437                         if (depth > 0)
 438                                 lfsck_object_put(env, obj);
 439                         return 1;
 440                 }
 441
 442                 /* .lustre doesn't contain "real" user objects, no need lfsck */
 443                 if (fid_is_dot_lustre(lfsck_dto2fid(obj))) {
 444                         if (depth > 0)
 445                                 lfsck_object_put(env, obj);
 446                         return 0;
 447                 }
 448
 449                 dt_read_lock(env, obj, MOR_TGT_CHILD);
 450                 if (unlikely(lfsck_is_dead_obj(obj))) {
 451                         dt_read_unlock(env, obj);
 452                         if (depth > 0)
 453                                 lfsck_object_put(env, obj);
 454                         return 0;
 455                 }
 456
 457                 rc = dt_xattr_get(env, obj,
 458                                   lfsck_buf_get(env, NULL, 0), XATTR_NAME_LINK,
 459                                   BYPASS_CAPA);
 460                 dt_read_unlock(env, obj);
 461                 if (rc >= 0) {
 462                         if (depth > 0)
 463                                 lfsck_object_put(env, obj);
 464                         return 1;
 465                 }
 466
 467                 if (rc < 0 && rc != -ENODATA) {
 468                         if (depth > 0)
 469                                 lfsck_object_put(env, obj);
 470                         return rc;
 471                 }
 472
 473                 rc = lfsck_parent_fid(env, obj, fid);
 474                 if (depth > 0)
 475                         lfsck_object_put(env, obj);
 476                 if (rc != 0)
 477                         return rc;
 478
 479                 if (unlikely(lu_fid_eq(fid, &lfsck->li_local_root_fid)))
 480                         return 0;
 481
 482                 obj = lfsck_object_find(env, lfsck, fid);
 483                 if (obj == NULL)
 484                         return 0;
 485                 else if (IS_ERR(obj))
 486                         return PTR_ERR(obj);
 487
 488                 if (!dt_object_exists(obj)) {
 489                         lfsck_object_put(env, obj);
 490                         return 0;
 491                 }
 492
 493                 /* Currently, only client visible directory can be remote. */
 494                 if (dt_object_remote(obj)) {
 495                         lfsck_object_put(env, obj);
 496                         return 1;
 497                 }
 498
 499                 depth++;
 500         }
 501         return 0;
 502 }
 503
 504 /* LFSCK wrap functions */
 505
 506 void lfsck_fail(const struct lu_env *env, struct lfsck_instance *lfsck,
 507                 bool new_checked)
 508 {
 509         struct lfsck_component *com;
 510
 511         cfs_list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
 512                 com->lc_ops->lfsck_fail(env, com, new_checked);
 513         }
 514 }
 515
 516 int lfsck_checkpoint(const struct lu_env *env, struct lfsck_instance *lfsck)
 517 {
 518         struct lfsck_component *com;
 519         int                     rc;
 520
 521         if (likely(cfs_time_beforeq(cfs_time_current(),
 522                                     lfsck->li_time_next_checkpoint)))
 523                 return 0;
 524
 525         lfsck_pos_fill(env, lfsck, &lfsck->li_pos_current, false);
 526         cfs_list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
 527                 rc = com->lc_ops->lfsck_checkpoint(env, com, false);
 528                 if (rc != 0)
 529                         return rc;;
 530         }
 531
 532         lfsck->li_time_last_checkpoint = cfs_time_current();
 533         lfsck->li_time_next_checkpoint = lfsck->li_time_last_checkpoint +
 534                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
 535         return 0;
 536 }
 537
 538 int lfsck_prep(const struct lu_env *env, struct lfsck_instance *lfsck)
 539 {
 540         struct dt_object       *obj     = NULL;
 541         struct lfsck_component *com;
 542         struct lfsck_component *next;
 543         struct lfsck_position  *pos     = NULL;
 544         const struct dt_it_ops *iops    =
 545                                 &lfsck->li_obj_oit->do_index_ops->dio_it;
 546         struct dt_it           *di;
 547         int                     rc;
 548         ENTRY;
 549
 550         LASSERT(lfsck->li_obj_dir == NULL);
 551         LASSERT(lfsck->li_di_dir == NULL);
 552
 553         lfsck->li_current_oit_processed = 0;
 554         cfs_list_for_each_entry_safe(com, next, &lfsck->li_list_scan, lc_link) {
 555                 com->lc_new_checked = 0;
 556                 if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
 557                         com->lc_journal = 0;
 558
 559                 rc = com->lc_ops->lfsck_prep(env, com);
 560                 if (rc != 0)
 561                         RETURN(rc);
 562
 563                 if ((pos == NULL) ||
 564                     (!lfsck_pos_is_zero(&com->lc_pos_start) &&
 565                      lfsck_pos_is_eq(pos, &com->lc_pos_start) > 0))
 566                         pos = &com->lc_pos_start;
 567         }
 568
 569         /* Init otable-based iterator. */
 570         if (pos == NULL) {
 571                 rc = iops->load(env, lfsck->li_di_oit, 0);
 572                 if (rc > 0) {
 573                         lfsck->li_oit_over = 1;
 574                         rc = 0;
 575                 }
 576
 577                 GOTO(out, rc);
 578         }
 579
 580         rc = iops->load(env, lfsck->li_di_oit, pos->lp_oit_cookie);
 581         if (rc < 0)
 582                 GOTO(out, rc);
 583         else if (rc > 0)
 584                 lfsck->li_oit_over = 1;
 585
 586         if (!lfsck->li_master || fid_is_zero(&pos->lp_dir_parent))
 587                 GOTO(out, rc = 0);
 588
 589         /* Find the directory for namespace-based traverse. */
 590         obj = lfsck_object_find(env, lfsck, &pos->lp_dir_parent);
 591         if (obj == NULL)
 592                 GOTO(out, rc = 0);
 593         else if (IS_ERR(obj))
 594                 RETURN(PTR_ERR(obj));
 595
 596         /* XXX: Currently, skip remote object, the consistency for
 597          *      remote object will be processed in LFSCK phase III. */
 598         if (!dt_object_exists(obj) || dt_object_remote(obj) ||
 599             unlikely(!S_ISDIR(lfsck_object_type(obj))))
 600                 GOTO(out, rc = 0);
 601
 602         if (unlikely(!dt_try_as_dir(env, obj)))
 603                 GOTO(out, rc = -ENOTDIR);
 604
 605         /* Init the namespace-based directory traverse. */
 606         iops = &obj->do_index_ops->dio_it;
 607         di = iops->init(env, obj, lfsck->li_args_dir, BYPASS_CAPA);
 608         if (IS_ERR(di))
 609                 GOTO(out, rc = PTR_ERR(di));
 610
 611         LASSERT(pos->lp_dir_cookie < MDS_DIR_END_OFF);
 612
 613         rc = iops->load(env, di, pos->lp_dir_cookie);
 614         if ((rc == 0) || (rc > 0 && pos->lp_dir_cookie > 0))
 615                 rc = iops->next(env, di);
 616         else if (rc > 0)
 617                 rc = 0;
 618
 619         if (rc != 0) {
 620                 iops->put(env, di);
 621                 iops->fini(env, di);
 622                 GOTO(out, rc);
 623         }
 624
 625         lfsck->li_obj_dir = lfsck_object_get(obj);
 626         lfsck->li_cookie_dir = iops->store(env, di);
 627         spin_lock(&lfsck->li_lock);
 628         lfsck->li_di_dir = di;
 629         spin_unlock(&lfsck->li_lock);
 630
 631         GOTO(out, rc = 0);
 632
 633 out:
 634         if (obj != NULL)
 635                 lfsck_object_put(env, obj);
 636
 637         if (rc < 0) {
 638                 cfs_list_for_each_entry_safe(com, next, &lfsck->li_list_scan,
 639                                              lc_link)
 640                         com->lc_ops->lfsck_post(env, com, rc, true);
 641
 642                 return rc;
 643         }
 644
 645         rc = 0;
 646         lfsck_pos_fill(env, lfsck, &lfsck->li_pos_current, true);
 647         cfs_list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
 648                 rc = com->lc_ops->lfsck_checkpoint(env, com, true);
 649                 if (rc != 0)
 650                         break;
 651         }
 652
 653         lfsck->li_time_last_checkpoint = cfs_time_current();
 654         lfsck->li_time_next_checkpoint = lfsck->li_time_last_checkpoint +
 655                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
 656         return rc;
 657 }
 658
 659 int lfsck_exec_oit(const struct lu_env *env, struct lfsck_instance *lfsck,
 660                    struct dt_object *obj)
 661 {
 662         struct lfsck_component *com;
 663         const struct dt_it_ops *iops;
 664         struct dt_it           *di;
 665         int                     rc;
 666         ENTRY;
 667
 668         LASSERT(lfsck->li_obj_dir == NULL);
 669
 670         cfs_list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
 671                 rc = com->lc_ops->lfsck_exec_oit(env, com, obj);
 672                 if (rc != 0)
 673                         RETURN(rc);
 674         }
 675
 676         rc = lfsck_needs_scan_dir(env, lfsck, obj);
 677         if (rc <= 0)
 678                 GOTO(out, rc);
 679
 680         if (unlikely(!dt_try_as_dir(env, obj)))
 681                 GOTO(out, rc = -ENOTDIR);
 682
 683         iops = &obj->do_index_ops->dio_it;
 684         di = iops->init(env, obj, lfsck->li_args_dir, BYPASS_CAPA);
 685         if (IS_ERR(di))
 686                 GOTO(out, rc = PTR_ERR(di));
 687
 688         rc = iops->load(env, di, 0);
 689         if (rc == 0)
 690                 rc = iops->next(env, di);
 691         else if (rc > 0)
 692                 rc = 0;
 693
 694         if (rc != 0) {
 695                 iops->put(env, di);
 696                 iops->fini(env, di);
 697                 GOTO(out, rc);
 698         }
 699
 700         lfsck->li_obj_dir = lfsck_object_get(obj);
 701         lfsck->li_cookie_dir = iops->store(env, di);
 702         spin_lock(&lfsck->li_lock);
 703         lfsck->li_di_dir = di;
 704         spin_unlock(&lfsck->li_lock);
 705
 706         GOTO(out, rc = 0);
 707
 708 out:
 709         if (rc < 0)
 710                 lfsck_fail(env, lfsck, false);
 711         return (rc > 0 ? 0 : rc);
 712 }
 713
 714 int lfsck_exec_dir(const struct lu_env *env, struct lfsck_instance *lfsck,
 715                    struct dt_object *obj, struct lu_dirent *ent)
 716 {
 717         struct lfsck_component *com;
 718         int                     rc;
 719
 720         cfs_list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
 721                 rc = com->lc_ops->lfsck_exec_dir(env, com, obj, ent);
 722                 if (rc != 0)
 723                         return rc;
 724         }
 725         return 0;
 726 }
 727
 728 int lfsck_post(const struct lu_env *env, struct lfsck_instance *lfsck,
 729                int result)
 730 {
 731         struct lfsck_component *com;
 732         struct lfsck_component *next;
 733         int                     rc;
 734
 735         lfsck_pos_fill(env, lfsck, &lfsck->li_pos_current, false);
 736         cfs_list_for_each_entry_safe(com, next, &lfsck->li_list_scan, lc_link) {
 737                 rc = com->lc_ops->lfsck_post(env, com, result, false);
 738                 if (rc != 0)
 739                         return rc;
 740         }
 741
 742         lfsck->li_time_last_checkpoint = cfs_time_current();
 743         lfsck->li_time_next_checkpoint = lfsck->li_time_last_checkpoint +
 744                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
 745         return result;
 746 }
 747
 748 int lfsck_double_scan(const struct lu_env *env, struct lfsck_instance *lfsck)
 749 {
 750         struct lfsck_component *com;
 751         struct lfsck_component *next;
 752         int                     rc;
 753
 754         cfs_list_for_each_entry_safe(com, next, &lfsck->li_list_double_scan,
 755                                      lc_link) {
 756                 if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
 757                         com->lc_journal = 0;
 758
 759                 rc = com->lc_ops->lfsck_double_scan(env, com);
 760                 if (rc != 0)
 761                         return rc;
 762         }
 763         return 0;
 764 }
 765
 766 /* external interfaces */
 767
 768 int lfsck_get_speed(struct dt_device *key, void *buf, int len)
 769 {
 770         struct lu_env           env;
 771         struct lfsck_instance  *lfsck;
 772         int                     rc;
 773         ENTRY;
 774
 775         lfsck = lfsck_instance_find(key, true, false);
 776         if (unlikely(lfsck == NULL))
 777                 RETURN(-ENODEV);
 778
 779         rc = lu_env_init(&env, LCT_MD_THREAD | LCT_DT_THREAD);
 780         if (rc != 0)
 781                 GOTO(out, rc);
 782
 783         rc = snprintf(buf, len, "%u\n", lfsck->li_bookmark_ram.lb_speed_limit);
 784         lu_env_fini(&env);
 785
 786         GOTO(out, rc);
 787
 788 out:
 789         lfsck_instance_put(&env, lfsck);
 790         return rc;
 791 }
 792 EXPORT_SYMBOL(lfsck_get_speed);
 793
 794 int lfsck_set_speed(struct dt_device *key, int val)
 795 {
 796         struct lu_env           env;
 797         struct lfsck_instance  *lfsck;
 798         int                     rc;
 799         ENTRY;
 800
 801         lfsck = lfsck_instance_find(key, true, false);
 802         if (unlikely(lfsck == NULL))
 803                 RETURN(-ENODEV);
 804
 805         rc = lu_env_init(&env, LCT_MD_THREAD | LCT_DT_THREAD);
 806         if (rc != 0)
 807                 GOTO(out, rc);
 808
 809         mutex_lock(&lfsck->li_mutex);
 810         __lfsck_set_speed(lfsck, val);
 811         rc = lfsck_bookmark_store(&env, lfsck);
 812         mutex_unlock(&lfsck->li_mutex);
 813         lu_env_fini(&env);
 814
 815         GOTO(out, rc);
 816
 817 out:
 818         lfsck_instance_put(&env, lfsck);
 819         return rc;
 820 }
 821 EXPORT_SYMBOL(lfsck_set_speed);
 822
 823 int lfsck_dump(struct dt_device *key, void *buf, int len, __u16 type)
 824 {
 825         struct lu_env           env;
 826         struct lfsck_instance  *lfsck;
 827         struct lfsck_component *com   = NULL;
 828         int                     rc;
 829         ENTRY;
 830
 831         lfsck = lfsck_instance_find(key, true, false);
 832         if (unlikely(lfsck == NULL))
 833                 RETURN(-ENODEV);
 834
 835         com = lfsck_component_find(lfsck, type);
 836         if (com == NULL)
 837                 GOTO(out, rc = -ENOTSUPP);
 838
 839         rc = lu_env_init(&env, LCT_MD_THREAD | LCT_DT_THREAD);
 840         if (rc != 0)
 841                 GOTO(out, rc);
 842
 843         rc = com->lc_ops->lfsck_dump(&env, com, buf, len);
 844         lu_env_fini(&env);
 845
 846         GOTO(out, rc);
 847
 848 out:
 849         if (com != NULL)
 850                 lfsck_component_put(&env, com);
 851         lfsck_instance_put(&env, lfsck);
 852         return rc;
 853 }
 854 EXPORT_SYMBOL(lfsck_dump);
 855
 856 int lfsck_start(const struct lu_env *env, struct dt_device *key,
 857                 struct lfsck_start_param *lsp)
 858 {
 859         struct lfsck_start     *start  = lsp->lsp_start;
 860         struct lfsck_instance  *lfsck;
 861         struct lfsck_bookmark  *bk;
 862         struct ptlrpc_thread   *thread;
 863         struct lfsck_component *com;
 864         struct l_wait_info      lwi    = { 0 };
 865         bool                    dirty  = false;
 866         long                    rc     = 0;
 867         __u16                   valid  = 0;
 868         __u16                   flags  = 0;
 869         ENTRY;
 870
 871         lfsck = lfsck_instance_find(key, true, false);
 872         if (unlikely(lfsck == NULL))
 873                 RETURN(-ENODEV);
 874
 875         /* start == NULL means auto trigger paused LFSCK. */
 876         if ((start == NULL) &&
 877             (cfs_list_empty(&lfsck->li_list_scan) ||
 878              OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NO_AUTO)))
 879                 GOTO(put, rc = 0);
 880
 881         bk = &lfsck->li_bookmark_ram;
 882         thread = &lfsck->li_thread;
 883         mutex_lock(&lfsck->li_mutex);
 884         spin_lock(&lfsck->li_lock);
 885         if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
 886                 spin_unlock(&lfsck->li_lock);
 887                 GOTO(out, rc = -EALREADY);
 888         }
 889
 890         spin_unlock(&lfsck->li_lock);
 891
 892         lfsck->li_namespace = lsp->lsp_namespace;
 893         lfsck->li_paused = 0;
 894         lfsck->li_oit_over = 0;
 895         lfsck->li_drop_dryrun = 0;
 896         lfsck->li_new_scanned = 0;
 897
 898         /* For auto trigger. */
 899         if (start == NULL)
 900                 goto trigger;
 901
 902         start->ls_version = bk->lb_version;
 903         if (start->ls_valid & LSV_SPEED_LIMIT) {
 904                 __lfsck_set_speed(lfsck, start->ls_speed_limit);
 905                 dirty = true;
 906         }
 907
 908         if (start->ls_valid & LSV_ERROR_HANDLE) {
 909                 valid |= DOIV_ERROR_HANDLE;
 910                 if (start->ls_flags & LPF_FAILOUT)
 911                         flags |= DOIF_FAILOUT;
 912
 913                 if ((start->ls_flags & LPF_FAILOUT) &&
 914                     !(bk->lb_param & LPF_FAILOUT)) {
 915                         bk->lb_param |= LPF_FAILOUT;
 916                         dirty = true;
 917                 } else if (!(start->ls_flags & LPF_FAILOUT) &&
 918                            (bk->lb_param & LPF_FAILOUT)) {
 919                         bk->lb_param &= ~LPF_FAILOUT;
 920                         dirty = true;
 921                 }
 922         }
 923
 924         if (start->ls_valid & LSV_DRYRUN) {
 925                 valid |= DOIV_DRYRUN;
 926                 if (start->ls_flags & LPF_DRYRUN)
 927                         flags |= DOIF_DRYRUN;
 928
 929                 if ((start->ls_flags & LPF_DRYRUN) &&
 930                     !(bk->lb_param & LPF_DRYRUN)) {
 931                         bk->lb_param |= LPF_DRYRUN;
 932                         dirty = true;
 933                 } else if (!(start->ls_flags & LPF_DRYRUN) &&
 934                            (bk->lb_param & LPF_DRYRUN)) {
 935                         bk->lb_param &= ~LPF_DRYRUN;
 936                         lfsck->li_drop_dryrun = 1;
 937                         dirty = true;
 938                 }
 939         }
 940
 941         if (dirty) {
 942                 rc = lfsck_bookmark_store(env, lfsck);
 943                 if (rc != 0)
 944                         GOTO(out, rc);
 945         }
 946
 947         if (start->ls_flags & LPF_RESET)
 948                 flags |= DOIF_RESET;
 949
 950         if (start->ls_active != 0) {
 951                 struct lfsck_component *next;
 952                 __u16 type = 1;
 953
 954                 if (start->ls_active == LFSCK_TYPES_ALL)
 955                         start->ls_active = LFSCK_TYPES_SUPPORTED;
 956
 957                 if (start->ls_active & ~LFSCK_TYPES_SUPPORTED) {
 958                         start->ls_active &= ~LFSCK_TYPES_SUPPORTED;
 959                         GOTO(out, rc = -ENOTSUPP);
 960                 }
 961
 962                 cfs_list_for_each_entry_safe(com, next,
 963                                              &lfsck->li_list_scan, lc_link) {
 964                         if (!(com->lc_type & start->ls_active)) {
 965                                 rc = com->lc_ops->lfsck_post(env, com, 0,
 966                                                              false);
 967                                 if (rc != 0)
 968                                         GOTO(out, rc);
 969                         }
 970                 }
 971
 972                 while (start->ls_active != 0) {
 973                         if (type & start->ls_active) {
 974                                 com = __lfsck_component_find(lfsck, type,
 975                                                         &lfsck->li_list_idle);
 976                                 if (com != NULL) {
 977                                         /* The component status will be updated
 978                                          * when its prep() is called later by
 979                                          * the LFSCK main engine. */
 980                                         cfs_list_del_init(&com->lc_link);
 981                                         cfs_list_add_tail(&com->lc_link,
 982                                                           &lfsck->li_list_scan);
 983                                 }
 984                                 start->ls_active &= ~type;
 985                         }
 986                         type <<= 1;
 987                 }
 988         }
 989
 990         cfs_list_for_each_entry(com, &lfsck->li_list_scan, lc_link) {
 991                 start->ls_active |= com->lc_type;
 992                 if (flags & DOIF_RESET) {
 993                         rc = com->lc_ops->lfsck_reset(env, com, false);
 994                         if (rc != 0)
 995                                 GOTO(out, rc);
 996                 }
 997         }
 998
 999 trigger:
1000         lfsck->li_args_dir = LUDA_64BITHASH | LUDA_VERIFY;
1001         if (bk->lb_param & LPF_DRYRUN) {
1002                 lfsck->li_args_dir |= LUDA_VERIFY_DRYRUN;
1003                 valid |= DOIV_DRYRUN;
1004                 flags |= DOIF_DRYRUN;
1005         }
1006
1007         if (bk->lb_param & LPF_FAILOUT) {
1008                 valid |= DOIV_ERROR_HANDLE;
1009                 flags |= DOIF_FAILOUT;
1010         }
1011
1012         if (!cfs_list_empty(&lfsck->li_list_scan))
1013                 flags |= DOIF_OUTUSED;
1014
1015         lfsck->li_args_oit = (flags << DT_OTABLE_IT_FLAGS_SHIFT) | valid;
1016         thread_set_flags(thread, 0);
1017         rc = PTR_ERR(kthread_run(lfsck_master_engine, lfsck, "lfsck"));
1018         if (IS_ERR_VALUE(rc)) {
1019                 CERROR("%s: cannot start LFSCK thread, rc = %ld\n",
1020                        lfsck_lfsck2name(lfsck), rc);
1021         } else {
1022                 rc = 0;
1023                 l_wait_event(thread->t_ctl_waitq,
1024                              thread_is_running(thread) ||
1025                              thread_is_stopped(thread),
1026                              &lwi);
1027         }
1028
1029         GOTO(out, rc);
1030
1031 out:
1032         mutex_unlock(&lfsck->li_mutex);
1033 put:
1034         lfsck_instance_put(env, lfsck);
1035         return (rc < 0 ? rc : 0);
1036 }
1037 EXPORT_SYMBOL(lfsck_start);
1038
1039 int lfsck_stop(const struct lu_env *env, struct dt_device *key, bool pause)
1040 {
1041         struct lfsck_instance   *lfsck;
1042         struct ptlrpc_thread    *thread;
1043         struct l_wait_info       lwi    = { 0 };
1044         ENTRY;
1045
1046         lfsck = lfsck_instance_find(key, true, false);
1047         if (unlikely(lfsck == NULL))
1048                 RETURN(-ENODEV);
1049
1050         thread = &lfsck->li_thread;
1051         mutex_lock(&lfsck->li_mutex);
1052         spin_lock(&lfsck->li_lock);
1053         if (thread_is_init(thread) || thread_is_stopped(thread)) {
1054                 spin_unlock(&lfsck->li_lock);
1055                 mutex_unlock(&lfsck->li_mutex);
1056                 lfsck_instance_put(env, lfsck);
1057                 RETURN(-EALREADY);
1058         }
1059
1060         if (pause)
1061                 lfsck->li_paused = 1;
1062         thread_set_flags(thread, SVC_STOPPING);
1063         spin_unlock(&lfsck->li_lock);
1064
1065         wake_up_all(&thread->t_ctl_waitq);
1066         l_wait_event(thread->t_ctl_waitq,
1067                      thread_is_stopped(thread),
1068                      &lwi);
1069         mutex_unlock(&lfsck->li_mutex);
1070         lfsck_instance_put(env, lfsck);
1071
1072         RETURN(0);
1073 }
1074 EXPORT_SYMBOL(lfsck_stop);
1075
1076 int lfsck_register(const struct lu_env *env, struct dt_device *key,
1077                    struct dt_device *next, bool master)
1078 {
1079         struct lfsck_instance   *lfsck;
1080         struct dt_object        *root  = NULL;
1081         struct dt_object        *obj;
1082         struct lu_fid           *fid   = &lfsck_env_info(env)->lti_fid;
1083         int                      rc;
1084         ENTRY;
1085
1086         lfsck = lfsck_instance_find(key, false, false);
1087         if (unlikely(lfsck != NULL))
1088                 RETURN(-EEXIST);
1089
1090         OBD_ALLOC_PTR(lfsck);
1091         if (lfsck == NULL)
1092                 RETURN(-ENOMEM);
1093
1094         mutex_init(&lfsck->li_mutex);
1095         spin_lock_init(&lfsck->li_lock);
1096         CFS_INIT_LIST_HEAD(&lfsck->li_link);
1097         CFS_INIT_LIST_HEAD(&lfsck->li_list_scan);
1098         CFS_INIT_LIST_HEAD(&lfsck->li_list_dir);
1099         CFS_INIT_LIST_HEAD(&lfsck->li_list_double_scan);
1100         CFS_INIT_LIST_HEAD(&lfsck->li_list_idle);
1101         atomic_set(&lfsck->li_ref, 1);
1102         init_waitqueue_head(&lfsck->li_thread.t_ctl_waitq);
1103         lfsck->li_next = next;
1104         lfsck->li_bottom = key;
1105
1106         fid->f_seq = FID_SEQ_LOCAL_NAME;
1107         fid->f_oid = 1;
1108         fid->f_ver = 0;
1109         rc = local_oid_storage_init(env, lfsck->li_bottom, fid, &lfsck->li_los);
1110         if (rc != 0)
1111                 GOTO(out, rc);
1112
1113         rc = dt_root_get(env, key, fid);
1114         if (rc != 0)
1115                 GOTO(out, rc);
1116
1117         root = dt_locate(env, lfsck->li_bottom, fid);
1118         if (IS_ERR(root))
1119                 GOTO(out, rc = PTR_ERR(root));
1120
1121         lfsck->li_local_root_fid = *fid;
1122         dt_try_as_dir(env, root);
1123         if (master) {
1124                 lfsck->li_master = 1;
1125                 if (lfsck_dev_idx(lfsck->li_bottom) == 0) {
1126                         rc = dt_lookup(env, root,
1127                                 (struct dt_rec *)(&lfsck->li_global_root_fid),
1128                                 (const struct dt_key *)"ROOT", BYPASS_CAPA);
1129                         if (rc != 0)
1130                                 GOTO(out, rc);
1131                 }
1132         }
1133
1134         fid->f_seq = FID_SEQ_LOCAL_FILE;
1135         fid->f_oid = OTABLE_IT_OID;
1136         fid->f_ver = 0;
1137         obj = dt_locate(env, lfsck->li_bottom, fid);
1138         if (IS_ERR(obj))
1139                 GOTO(out, rc = PTR_ERR(obj));
1140
1141         lfsck->li_obj_oit = obj;
1142         rc = obj->do_ops->do_index_try(env, obj, &dt_otable_features);
1143         if (rc != 0) {
1144                 if (rc == -ENOTSUPP)
1145                         GOTO(add, rc = 0);
1146
1147                 GOTO(out, rc);
1148         }
1149
1150         rc = lfsck_bookmark_setup(env, lfsck);
1151         if (rc != 0)
1152                 GOTO(out, rc);
1153
1154         if (master) {
1155                 rc = lfsck_namespace_setup(env, lfsck);
1156                 if (rc < 0)
1157                         GOTO(out, rc);
1158         }
1159
1160         /* XXX: more LFSCK components initialization to be added here. */
1161
1162 add:
1163         rc = lfsck_instance_add(lfsck);
1164 out:
1165         if (root != NULL && !IS_ERR(root))
1166                 lu_object_put(env, &root->do_lu);
1167         if (rc != 0)
1168                 lfsck_instance_cleanup(env, lfsck);
1169         return rc;
1170 }
1171 EXPORT_SYMBOL(lfsck_register);
1172
1173 void lfsck_degister(const struct lu_env *env, struct dt_device *key)
1174 {
1175         struct lfsck_instance *lfsck;
1176
1177         lfsck = lfsck_instance_find(key, false, true);
1178         if (lfsck != NULL)
1179                 lfsck_instance_put(env, lfsck);
1180 }
1181 EXPORT_SYMBOL(lfsck_degister);
1182
1183 static int __init lfsck_init(void)
1184 {
1185         int rc;
1186
1187         lfsck_key_init_generic(&lfsck_thread_key, NULL);
1188         rc = lu_context_key_register(&lfsck_thread_key);
1189         return rc;
1190 }
1191
1192 static void __exit lfsck_exit(void)
1193 {
1194         LASSERT(cfs_list_empty(&lfsck_instance_list));
1195
1196         lu_context_key_degister(&lfsck_thread_key);
1197 }
1198
1199 MODULE_AUTHOR("Intel Corporation <http://www.intel.com/>");
1200 MODULE_DESCRIPTION("LFSCK");
1201 MODULE_LICENSE("GPL");
1202
1203 cfs_module(lfsck, LUSTRE_VERSION_STRING, lfsck_init, lfsck_exit);