lustre/obdclass/scrub.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, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2017, Intel Corporation.
  24  */
  25 /*
  26  * lustre/obdclass/scrub.c
  27  *
  28  * The OI scrub is used for checking and (re)building Object Index files
  29  * that are usually backend special. Here are some general scrub related
  30  * functions that can be shared by different backends for OI scrub.
  31  *
  32  * Author: Fan Yong <fan.yong@intel.com>
  33  */
  34
  35 #define DEBUG_SUBSYSTEM S_LFSCK
  36
  37 #include <linux/kthread.h>
  38 #include <lustre_scrub.h>
  39 #include <lustre_lib.h>
  40 #include <lustre_fid.h>
  41
  42 static inline struct dt_device *scrub_obj2dev(struct dt_object *obj)
  43 {
  44         return container_of0(obj->do_lu.lo_dev, struct dt_device, dd_lu_dev);
  45 }
  46
  47 static void scrub_file_to_cpu(struct scrub_file *des, struct scrub_file *src)
  48 {
  49         uuid_copy(&des->sf_uuid, &src->sf_uuid);
  50         des->sf_flags   = le64_to_cpu(src->sf_flags);
  51         des->sf_magic   = le32_to_cpu(src->sf_magic);
  52         des->sf_status  = le16_to_cpu(src->sf_status);
  53         des->sf_param   = le16_to_cpu(src->sf_param);
  54         des->sf_time_last_complete      =
  55                                 le64_to_cpu(src->sf_time_last_complete);
  56         des->sf_time_latest_start       =
  57                                 le64_to_cpu(src->sf_time_latest_start);
  58         des->sf_time_last_checkpoint    =
  59                                 le64_to_cpu(src->sf_time_last_checkpoint);
  60         des->sf_pos_latest_start        =
  61                                 le64_to_cpu(src->sf_pos_latest_start);
  62         des->sf_pos_last_checkpoint     =
  63                                 le64_to_cpu(src->sf_pos_last_checkpoint);
  64         des->sf_pos_first_inconsistent  =
  65                                 le64_to_cpu(src->sf_pos_first_inconsistent);
  66         des->sf_items_checked           =
  67                                 le64_to_cpu(src->sf_items_checked);
  68         des->sf_items_updated           =
  69                                 le64_to_cpu(src->sf_items_updated);
  70         des->sf_items_failed            =
  71                                 le64_to_cpu(src->sf_items_failed);
  72         des->sf_items_updated_prior     =
  73                                 le64_to_cpu(src->sf_items_updated_prior);
  74         des->sf_run_time        = le32_to_cpu(src->sf_run_time);
  75         des->sf_success_count   = le32_to_cpu(src->sf_success_count);
  76         des->sf_oi_count        = le16_to_cpu(src->sf_oi_count);
  77         des->sf_internal_flags  = le16_to_cpu(src->sf_internal_flags);
  78         memcpy(des->sf_oi_bitmap, src->sf_oi_bitmap, SCRUB_OI_BITMAP_SIZE);
  79 }
  80
  81 static void scrub_file_to_le(struct scrub_file *des, struct scrub_file *src)
  82 {
  83         uuid_copy(&des->sf_uuid, &src->sf_uuid);
  84         des->sf_flags   = cpu_to_le64(src->sf_flags);
  85         des->sf_magic   = cpu_to_le32(src->sf_magic);
  86         des->sf_status  = cpu_to_le16(src->sf_status);
  87         des->sf_param   = cpu_to_le16(src->sf_param);
  88         des->sf_time_last_complete      =
  89                                 cpu_to_le64(src->sf_time_last_complete);
  90         des->sf_time_latest_start       =
  91                                 cpu_to_le64(src->sf_time_latest_start);
  92         des->sf_time_last_checkpoint    =
  93                                 cpu_to_le64(src->sf_time_last_checkpoint);
  94         des->sf_pos_latest_start        =
  95                                 cpu_to_le64(src->sf_pos_latest_start);
  96         des->sf_pos_last_checkpoint     =
  97                                 cpu_to_le64(src->sf_pos_last_checkpoint);
  98         des->sf_pos_first_inconsistent  =
  99                                 cpu_to_le64(src->sf_pos_first_inconsistent);
 100         des->sf_items_checked           =
 101                                 cpu_to_le64(src->sf_items_checked);
 102         des->sf_items_updated           =
 103                                 cpu_to_le64(src->sf_items_updated);
 104         des->sf_items_failed            =
 105                                 cpu_to_le64(src->sf_items_failed);
 106         des->sf_items_updated_prior     =
 107                                 cpu_to_le64(src->sf_items_updated_prior);
 108         des->sf_run_time        = cpu_to_le32(src->sf_run_time);
 109         des->sf_success_count   = cpu_to_le32(src->sf_success_count);
 110         des->sf_oi_count        = cpu_to_le16(src->sf_oi_count);
 111         des->sf_internal_flags  = cpu_to_le16(src->sf_internal_flags);
 112         memcpy(des->sf_oi_bitmap, src->sf_oi_bitmap, SCRUB_OI_BITMAP_SIZE);
 113 }
 114
 115 void scrub_file_init(struct lustre_scrub *scrub, uuid_t uuid)
 116 {
 117         struct scrub_file *sf = &scrub->os_file;
 118
 119         memset(sf, 0, sizeof(*sf));
 120         uuid_copy(&sf->sf_uuid, &uuid);
 121         sf->sf_magic = SCRUB_MAGIC_V1;
 122         sf->sf_status = SS_INIT;
 123 }
 124 EXPORT_SYMBOL(scrub_file_init);
 125
 126 void scrub_file_reset(struct lustre_scrub *scrub, uuid_t uuid, u64 flags)
 127 {
 128         struct scrub_file *sf = &scrub->os_file;
 129
 130         CDEBUG(D_LFSCK, "%s: reset OI scrub file, old flags = "
 131                "%#llx, add flags = %#llx\n",
 132                scrub->os_name, sf->sf_flags, flags);
 133
 134         uuid_copy(&sf->sf_uuid, &uuid);
 135         sf->sf_status = SS_INIT;
 136         sf->sf_flags |= flags;
 137         sf->sf_flags &= ~SF_AUTO;
 138         sf->sf_run_time = 0;
 139         sf->sf_time_latest_start = 0;
 140         sf->sf_time_last_checkpoint = 0;
 141         sf->sf_pos_latest_start = 0;
 142         sf->sf_pos_last_checkpoint = 0;
 143         sf->sf_pos_first_inconsistent = 0;
 144         sf->sf_items_checked = 0;
 145         sf->sf_items_updated = 0;
 146         sf->sf_items_failed = 0;
 147         sf->sf_items_noscrub = 0;
 148         sf->sf_items_igif = 0;
 149         if (!scrub->os_in_join)
 150                 sf->sf_items_updated_prior = 0;
 151 }
 152 EXPORT_SYMBOL(scrub_file_reset);
 153
 154 int scrub_file_load(const struct lu_env *env, struct lustre_scrub *scrub)
 155 {
 156         struct scrub_file *sf = &scrub->os_file;
 157         struct lu_buf buf = {
 158                 .lb_buf = &scrub->os_file_disk,
 159                 .lb_len = sizeof(scrub->os_file_disk)
 160         };
 161         loff_t pos = 0;
 162         int rc;
 163
 164         rc = dt_read(env, scrub->os_obj, &buf, &pos);
 165         /* failure */
 166         if (rc < 0) {
 167                 CERROR("%s: fail to load scrub file: rc = %d\n",
 168                        scrub->os_name, rc);
 169                 return rc;
 170         }
 171
 172         /* empty */
 173         if (!rc)
 174                 return -ENOENT;
 175
 176         /* corrupted */
 177         if (rc < buf.lb_len) {
 178                 CDEBUG(D_LFSCK, "%s: fail to load scrub file, "
 179                        "expected = %d: rc = %d\n",
 180                        scrub->os_name, (int)buf.lb_len, rc);
 181                 return -EFAULT;
 182         }
 183
 184         scrub_file_to_cpu(sf, &scrub->os_file_disk);
 185         if (sf->sf_magic != SCRUB_MAGIC_V1) {
 186                 CDEBUG(D_LFSCK, "%s: invalid scrub magic 0x%x != 0x%x\n",
 187                        scrub->os_name, sf->sf_magic, SCRUB_MAGIC_V1);
 188                 return -EFAULT;
 189         }
 190
 191         return 0;
 192 }
 193 EXPORT_SYMBOL(scrub_file_load);
 194
 195 int scrub_file_store(const struct lu_env *env, struct lustre_scrub *scrub)
 196 {
 197         struct scrub_file *sf = &scrub->os_file_disk;
 198         struct dt_object *obj = scrub->os_obj;
 199         struct dt_device *dev = scrub_obj2dev(obj);
 200         struct lu_buf buf = {
 201                 .lb_buf = sf,
 202                 .lb_len = sizeof(*sf)
 203         };
 204         struct thandle *th;
 205         loff_t pos = 0;
 206         int rc;
 207         ENTRY;
 208
 209         /* Skip store under rdonly mode. */
 210         if (dev->dd_rdonly)
 211                 RETURN(0);
 212
 213         scrub_file_to_le(sf, &scrub->os_file);
 214         th = dt_trans_create(env, dev);
 215         if (IS_ERR(th))
 216                 GOTO(log, rc = PTR_ERR(th));
 217
 218         rc = dt_declare_record_write(env, obj, &buf, pos, th);
 219         if (rc)
 220                 GOTO(stop, rc);
 221
 222         rc = dt_trans_start_local(env, dev, th);
 223         if (rc)
 224                 GOTO(stop, rc);
 225
 226         rc = dt_record_write(env, obj, &buf, &pos, th);
 227
 228         GOTO(stop, rc);
 229
 230 stop:
 231         dt_trans_stop(env, dev, th);
 232
 233 log:
 234         if (rc)
 235                 CERROR("%s: store scrub file: rc = %d\n",
 236                        scrub->os_name, rc);
 237         else
 238                 CDEBUG(D_LFSCK, "%s: store scrub file: rc = %d\n",
 239                        scrub->os_name, rc);
 240
 241         scrub->os_time_last_checkpoint = ktime_get_seconds();
 242         scrub->os_time_next_checkpoint = scrub->os_time_last_checkpoint +
 243                                          SCRUB_CHECKPOINT_INTERVAL;
 244         return rc;
 245 }
 246 EXPORT_SYMBOL(scrub_file_store);
 247
 248 int scrub_checkpoint(const struct lu_env *env, struct lustre_scrub *scrub)
 249 {
 250         struct scrub_file *sf = &scrub->os_file;
 251         time64_t now = ktime_get_seconds();
 252         int rc;
 253
 254         if (likely(now < scrub->os_time_next_checkpoint ||
 255                    scrub->os_new_checked == 0))
 256                 return 0;
 257
 258         CDEBUG(D_LFSCK, "%s: OI scrub checkpoint at pos %llu\n",
 259                scrub->os_name, scrub->os_pos_current);
 260
 261         down_write(&scrub->os_rwsem);
 262         sf->sf_items_checked += scrub->os_new_checked;
 263         scrub->os_new_checked = 0;
 264         sf->sf_pos_last_checkpoint = scrub->os_pos_current;
 265         sf->sf_time_last_checkpoint = ktime_get_real_seconds();
 266         sf->sf_run_time += now - scrub->os_time_last_checkpoint;
 267         rc = scrub_file_store(env, scrub);
 268         up_write(&scrub->os_rwsem);
 269
 270         return rc;
 271 }
 272 EXPORT_SYMBOL(scrub_checkpoint);
 273
 274 int scrub_start(int (*threadfn)(void *data), struct lustre_scrub *scrub,
 275                 void *data, __u32 flags)
 276 {
 277         struct ptlrpc_thread *thread = &scrub->os_thread;
 278         struct task_struct *task;
 279         int rc;
 280         ENTRY;
 281
 282 again:
 283         /* os_lock: sync status between stop and scrub thread */
 284         spin_lock(&scrub->os_lock);
 285         if (thread_is_running(thread)) {
 286                 spin_unlock(&scrub->os_lock);
 287                 RETURN(-EALREADY);
 288         }
 289
 290         if (unlikely(thread_is_stopping(thread))) {
 291                 spin_unlock(&scrub->os_lock);
 292                 wait_event_idle(thread->t_ctl_waitq,
 293                                 thread_is_stopped(thread));
 294                 goto again;
 295         }
 296         spin_unlock(&scrub->os_lock);
 297
 298         if (scrub->os_file.sf_status == SS_COMPLETED) {
 299                 if (!(flags & SS_SET_FAILOUT))
 300                         flags |= SS_CLEAR_FAILOUT;
 301
 302                 if (!(flags & SS_SET_DRYRUN))
 303                         flags |= SS_CLEAR_DRYRUN;
 304
 305                 flags |= SS_RESET;
 306         }
 307
 308         scrub->os_start_flags = flags;
 309         thread_set_flags(thread, 0);
 310         task = kthread_run(threadfn, data, "OI_scrub");
 311         if (IS_ERR(task)) {
 312                 rc = PTR_ERR(task);
 313                 CERROR("%s: cannot start iteration thread: rc = %d\n",
 314                        scrub->os_name, rc);
 315                 RETURN(rc);
 316         }
 317
 318         wait_event_idle(thread->t_ctl_waitq,
 319                         thread_is_running(thread) || thread_is_stopped(thread));
 320
 321         RETURN(0);
 322 }
 323 EXPORT_SYMBOL(scrub_start);
 324
 325 void scrub_stop(struct lustre_scrub *scrub)
 326 {
 327         struct ptlrpc_thread *thread = &scrub->os_thread;
 328
 329         /* os_lock: sync status between stop and scrub thread */
 330         spin_lock(&scrub->os_lock);
 331         if (!thread_is_init(thread) && !thread_is_stopped(thread)) {
 332                 thread_set_flags(thread, SVC_STOPPING);
 333                 spin_unlock(&scrub->os_lock);
 334                 wake_up_all(&thread->t_ctl_waitq);
 335                 wait_event_idle(thread->t_ctl_waitq,
 336                                 thread_is_stopped(thread));
 337                 /* Do not skip the last lock/unlock, which can guarantee that
 338                  * the caller cannot return until the OI scrub thread exit. */
 339                 spin_lock(&scrub->os_lock);
 340         }
 341         spin_unlock(&scrub->os_lock);
 342 }
 343 EXPORT_SYMBOL(scrub_stop);
 344
 345 const char *scrub_status_names[] = {
 346         "init",
 347         "scanning",
 348         "completed",
 349         "failed",
 350         "stopped",
 351         "paused",
 352         "crashed",
 353         NULL
 354 };
 355
 356 const char *scrub_flags_names[] = {
 357         "recreated",
 358         "inconsistent",
 359         "auto",
 360         "upgrade",
 361         NULL
 362 };
 363
 364 const char *scrub_param_names[] = {
 365         "failout",
 366         "dryrun",
 367         NULL
 368 };
 369
 370 static void scrub_bits_dump(struct seq_file *m, int bits, const char *names[],
 371                             const char *prefix)
 372 {
 373         int flag;
 374         int i;
 375
 376         seq_printf(m, "%s:%c", prefix, bits != 0 ? ' ' : '\n');
 377
 378         for (i = 0, flag = 1; bits != 0; i++, flag = 1 << i) {
 379                 if (flag & bits) {
 380                         bits &= ~flag;
 381                         seq_printf(m, "%s%c", names[i],
 382                                    bits != 0 ? ',' : '\n');
 383                 }
 384         }
 385 }
 386
 387 static void scrub_time_dump(struct seq_file *m, time64_t time,
 388                             const char *prefix)
 389 {
 390         if (time != 0)
 391                 seq_printf(m, "%s: %llu seconds\n", prefix,
 392                            ktime_get_real_seconds() - time);
 393         else
 394                 seq_printf(m, "%s: N/A\n", prefix);
 395 }
 396
 397 static void scrub_pos_dump(struct seq_file *m, __u64 pos, const char *prefix)
 398 {
 399         if (pos != 0)
 400                 seq_printf(m, "%s: %llu\n", prefix, pos);
 401         else
 402                 seq_printf(m, "%s: N/A\n", prefix);
 403 }
 404
 405 void scrub_dump(struct seq_file *m, struct lustre_scrub *scrub)
 406 {
 407         struct scrub_file *sf = &scrub->os_file;
 408         u64 checked;
 409         s64 speed;
 410
 411         down_read(&scrub->os_rwsem);
 412         seq_printf(m, "name: OI_scrub\n"
 413                    "magic: 0x%x\n"
 414                    "oi_files: %d\n"
 415                    "status: %s\n",
 416                    sf->sf_magic, (int)sf->sf_oi_count,
 417                    scrub_status_names[sf->sf_status]);
 418
 419         scrub_bits_dump(m, sf->sf_flags, scrub_flags_names, "flags");
 420
 421         scrub_bits_dump(m, sf->sf_param, scrub_param_names, "param");
 422
 423         scrub_time_dump(m, sf->sf_time_last_complete,
 424                         "time_since_last_completed");
 425
 426         scrub_time_dump(m, sf->sf_time_latest_start,
 427                         "time_since_latest_start");
 428
 429         scrub_time_dump(m, sf->sf_time_last_checkpoint,
 430                         "time_since_last_checkpoint");
 431
 432         scrub_pos_dump(m, sf->sf_pos_latest_start,
 433                         "latest_start_position");
 434
 435         scrub_pos_dump(m, sf->sf_pos_last_checkpoint,
 436                         "last_checkpoint_position");
 437
 438         scrub_pos_dump(m, sf->sf_pos_first_inconsistent,
 439                         "first_failure_position");
 440
 441         checked = sf->sf_items_checked + scrub->os_new_checked;
 442         seq_printf(m, "checked: %llu\n"
 443                    "%s: %llu\n"
 444                    "failed: %llu\n"
 445                    "prior_%s: %llu\n"
 446                    "noscrub: %llu\n"
 447                    "igif: %llu\n"
 448                    "success_count: %u\n",
 449                    checked,
 450                    sf->sf_param & SP_DRYRUN ? "inconsistent" : "updated",
 451                    sf->sf_items_updated, sf->sf_items_failed,
 452                    sf->sf_param & SP_DRYRUN ? "inconsistent" : "updated",
 453                    sf->sf_items_updated_prior, sf->sf_items_noscrub,
 454                    sf->sf_items_igif, sf->sf_success_count);
 455
 456         speed = checked;
 457         if (thread_is_running(&scrub->os_thread)) {
 458                 s64 new_checked = scrub->os_new_checked;
 459                 time64_t duration;
 460                 time64_t rtime;
 461
 462                 /* Since the time resolution is in seconds for new system
 463                  * or small devices it ismore likely that duration will be
 464                  * zero which will lead to inaccurate results.
 465                  */
 466                 duration = ktime_get_seconds() -
 467                            scrub->os_time_last_checkpoint;
 468                 if (duration != 0)
 469                         new_checked = div_s64(new_checked, duration);
 470
 471                 rtime = sf->sf_run_time + duration;
 472                 if (rtime != 0)
 473                         speed = div_s64(speed, rtime);
 474
 475                 seq_printf(m, "run_time: %lld seconds\n"
 476                            "average_speed: %lld objects/sec\n"
 477                            "real_time_speed: %lld objects/sec\n"
 478                            "current_position: %llu\n"
 479                            "scrub_in_prior: %s\n"
 480                            "scrub_full_speed: %s\n"
 481                            "partial_scan: %s\n",
 482                            rtime, speed, new_checked,
 483                            scrub->os_pos_current,
 484                            scrub->os_in_prior ? "yes" : "no",
 485                            scrub->os_full_speed ? "yes" : "no",
 486                            scrub->os_partial_scan ? "yes" : "no");
 487         } else {
 488                 if (sf->sf_run_time != 0)
 489                         speed = div_s64(speed, sf->sf_run_time);
 490                 seq_printf(m, "run_time: %ld seconds\n"
 491                            "average_speed: %lld objects/sec\n"
 492                            "real_time_speed: N/A\n"
 493                            "current_position: N/A\n",
 494                            sf->sf_run_time, speed);
 495         }
 496
 497         up_read(&scrub->os_rwsem);
 498 }
 499 EXPORT_SYMBOL(scrub_dump);
 500
 501 int lustre_liru_new(struct list_head *head, const struct lu_fid *pfid,
 502                     const struct lu_fid *cfid, __u64 child,
 503                     const char *name, int namelen)
 504 {
 505         struct lustre_index_restore_unit *liru;
 506         int len = sizeof(*liru) + namelen + 1;
 507
 508         OBD_ALLOC(liru, len);
 509         if (!liru)
 510                 return -ENOMEM;
 511
 512         INIT_LIST_HEAD(&liru->liru_link);
 513         liru->liru_pfid = *pfid;
 514         liru->liru_cfid = *cfid;
 515         liru->liru_clid = child;
 516         liru->liru_len = len;
 517         memcpy(liru->liru_name, name, namelen);
 518         liru->liru_name[namelen] = 0;
 519         list_add_tail(&liru->liru_link, head);
 520
 521         return 0;
 522 }
 523 EXPORT_SYMBOL(lustre_liru_new);
 524
 525 int lustre_index_register(struct dt_device *dev, const char *devname,
 526                           struct list_head *head, spinlock_t *lock, int *guard,
 527                           const struct lu_fid *fid,
 528                           __u32 keysize, __u32 recsize)
 529 {
 530         struct lustre_index_backup_unit *libu, *pos;
 531         int rc = 0;
 532         ENTRY;
 533
 534         if (dev->dd_rdonly || *guard)
 535                 RETURN(1);
 536
 537         OBD_ALLOC_PTR(libu);
 538         if (!libu)
 539                 RETURN(-ENOMEM);
 540
 541         INIT_LIST_HEAD(&libu->libu_link);
 542         libu->libu_keysize = keysize;
 543         libu->libu_recsize = recsize;
 544         libu->libu_fid = *fid;
 545
 546         spin_lock(lock);
 547         if (unlikely(*guard)) {
 548                 spin_unlock(lock);
 549                 OBD_FREE_PTR(libu);
 550
 551                 RETURN(1);
 552         }
 553
 554         list_for_each_entry_reverse(pos, head, libu_link) {
 555                 rc = lu_fid_cmp(&pos->libu_fid, fid);
 556                 if (rc < 0) {
 557                         list_add(&libu->libu_link, &pos->libu_link);
 558                         spin_unlock(lock);
 559
 560                         RETURN(0);
 561                 }
 562
 563                 if (!rc) {
 564                         /* Registered already. But the former registered one
 565                          * has different keysize/recsize. It may because that
 566                          * the former values are from disk and corrupted, then
 567                          * replace it with new values. */
 568                         if (unlikely(keysize != pos->libu_keysize ||
 569                                      recsize != pos->libu_recsize)) {
 570                                 CWARN("%s: the index "DFID" has registered "
 571                                       "with %u/%u, may be invalid, replace "
 572                                       "with %u/%u\n",
 573                                       devname, PFID(fid), pos->libu_keysize,
 574                                       pos->libu_recsize, keysize, recsize);
 575
 576                                 pos->libu_keysize = keysize;
 577                                 pos->libu_recsize = recsize;
 578                         } else {
 579                                 rc = 1;
 580                         }
 581
 582                         spin_unlock(lock);
 583                         OBD_FREE_PTR(libu);
 584
 585                         RETURN(rc);
 586                 }
 587         }
 588
 589         list_add(&libu->libu_link, head);
 590         spin_unlock(lock);
 591
 592         RETURN(0);
 593 }
 594 EXPORT_SYMBOL(lustre_index_register);
 595
 596 static void lustre_index_degister(struct list_head *head, spinlock_t *lock,
 597                                   const struct lu_fid *fid)
 598 {
 599         struct lustre_index_backup_unit *libu;
 600         int rc = -ENOENT;
 601
 602         spin_lock(lock);
 603         list_for_each_entry_reverse(libu, head, libu_link) {
 604                 rc = lu_fid_cmp(&libu->libu_fid, fid);
 605                 /* NOT registered. */
 606                 if (rc < 0)
 607                         break;
 608
 609                 if (!rc) {
 610                         list_del(&libu->libu_link);
 611                         break;
 612                 }
 613         }
 614         spin_unlock(lock);
 615
 616         if (!rc)
 617                 OBD_FREE_PTR(libu);
 618 }
 619
 620 static void
 621 lustre_index_backup_make_header(struct lustre_index_backup_header *header,
 622                                 __u32 keysize, __u32 recsize,
 623                                 const struct lu_fid *fid, __u32 count)
 624 {
 625         memset(header, 0, sizeof(*header));
 626         header->libh_magic = cpu_to_le32(INDEX_BACKUP_MAGIC_V1);
 627         header->libh_count = cpu_to_le32(count);
 628         header->libh_keysize = cpu_to_le32(keysize);
 629         header->libh_recsize = cpu_to_le32(recsize);
 630         fid_cpu_to_le(&header->libh_owner, fid);
 631 }
 632
 633 static int lustre_index_backup_body(const struct lu_env *env,
 634                                     struct dt_object *obj, loff_t *pos,
 635                                     void *buf, int bufsize)
 636 {
 637         struct dt_device *dev = lu2dt_dev(obj->do_lu.lo_dev);
 638         struct thandle *th;
 639         struct lu_buf lbuf = {
 640                 .lb_buf = buf,
 641                 .lb_len = bufsize
 642         };
 643         int rc;
 644         ENTRY;
 645
 646         th = dt_trans_create(env, dev);
 647         if (IS_ERR(th))
 648                 RETURN(PTR_ERR(th));
 649
 650         rc = dt_declare_record_write(env, obj, &lbuf, *pos, th);
 651         if (rc)
 652                 GOTO(stop, rc);
 653
 654         rc = dt_trans_start_local(env, dev, th);
 655         if (rc)
 656                 GOTO(stop, rc);
 657
 658         rc = dt_record_write(env, obj, &lbuf, pos, th);
 659
 660         GOTO(stop, rc);
 661
 662 stop:
 663         dt_trans_stop(env, dev, th);
 664         return rc;
 665 }
 666
 667 static int lustre_index_backup_header(const struct lu_env *env,
 668                                       struct dt_object *obj,
 669                                       const struct lu_fid *tgt_fid,
 670                                       __u32 keysize, __u32 recsize,
 671                                       void *buf, int bufsize, int count)
 672 {
 673         struct dt_device *dev = lu2dt_dev(obj->do_lu.lo_dev);
 674         struct lustre_index_backup_header *header = buf;
 675         struct lu_attr *la = buf;
 676         struct thandle *th;
 677         struct lu_buf lbuf = {
 678                 .lb_buf = header,
 679                 .lb_len = sizeof(*header)
 680         };
 681         loff_t size = sizeof(*header) + (keysize + recsize) * count;
 682         loff_t pos = 0;
 683         int rc;
 684         bool punch = false;
 685         ENTRY;
 686
 687         LASSERT(sizeof(*la) <= bufsize);
 688         LASSERT(sizeof(*header) <= bufsize);
 689
 690         rc = dt_attr_get(env, obj, la);
 691         if (rc)
 692                 RETURN(rc);
 693
 694         if (la->la_size > size)
 695                 punch = true;
 696
 697         lustre_index_backup_make_header(header, keysize, recsize,
 698                                         tgt_fid, count);
 699         th = dt_trans_create(env, dev);
 700         if (IS_ERR(th))
 701                 RETURN(PTR_ERR(th));
 702
 703         rc = dt_declare_record_write(env, obj, &lbuf, pos, th);
 704         if (rc)
 705                 GOTO(stop, rc);
 706
 707         if (punch) {
 708                 rc = dt_declare_punch(env, obj, size, OBD_OBJECT_EOF, th);
 709                 if (rc)
 710                         GOTO(stop, rc);
 711         }
 712
 713         rc = dt_trans_start_local(env, dev, th);
 714         if (rc)
 715                 GOTO(stop, rc);
 716
 717         rc = dt_record_write(env, obj, &lbuf, &pos, th);
 718         if (!rc && punch)
 719                 rc = dt_punch(env, obj, size, OBD_OBJECT_EOF, th);
 720
 721         GOTO(stop, rc);
 722
 723 stop:
 724         dt_trans_stop(env, dev, th);
 725         return rc;
 726 }
 727
 728 static int lustre_index_update_lma(const struct lu_env *env,
 729                                    struct dt_object *obj,
 730                                    void *buf, int bufsize)
 731 {
 732         struct dt_device *dev = lu2dt_dev(obj->do_lu.lo_dev);
 733         struct lustre_mdt_attrs *lma = buf;
 734         struct lu_buf lbuf = {
 735                 .lb_buf = lma,
 736                 .lb_len = sizeof(struct lustre_ost_attrs)
 737         };
 738         struct thandle *th;
 739         int fl = LU_XATTR_REPLACE;
 740         int rc;
 741         ENTRY;
 742
 743         LASSERT(bufsize >= lbuf.lb_len);
 744
 745         rc = dt_xattr_get(env, obj, &lbuf, XATTR_NAME_LMA);
 746         if (unlikely(rc == -ENODATA)) {
 747                 fl = LU_XATTR_CREATE;
 748                 lustre_lma_init(lma, lu_object_fid(&obj->do_lu),
 749                                 LMAC_IDX_BACKUP, 0);
 750                 rc = sizeof(*lma);
 751         } else if (rc < sizeof(*lma)) {
 752                 RETURN(rc < 0 ? rc : -EFAULT);
 753         } else {
 754                 lustre_lma_swab(lma);
 755                 if (lma->lma_compat & LMAC_IDX_BACKUP)
 756                         RETURN(0);
 757
 758                 lma->lma_compat |= LMAC_IDX_BACKUP;
 759         }
 760
 761         lustre_lma_swab(lma);
 762         lbuf.lb_len = rc;
 763         th = dt_trans_create(env, dev);
 764         if (IS_ERR(th))
 765                 RETURN(rc);
 766
 767         rc = dt_declare_xattr_set(env, obj, &lbuf, XATTR_NAME_LMA, fl, th);
 768         if (rc)
 769                 GOTO(stop, rc);
 770
 771         rc = dt_trans_start_local(env, dev, th);
 772         if (rc)
 773                 GOTO(stop, rc);
 774
 775         rc = dt_xattr_set(env, obj, &lbuf, XATTR_NAME_LMA, fl, th);
 776
 777         GOTO(stop, rc);
 778
 779 stop:
 780         dt_trans_stop(env, dev, th);
 781         return rc;
 782 }
 783
 784 static int lustre_index_backup_one(const struct lu_env *env,
 785                                    struct local_oid_storage *los,
 786                                    struct dt_object *parent,
 787                                    struct lustre_index_backup_unit *libu,
 788                                    char *buf, int bufsize)
 789 {
 790         struct dt_device *dev = scrub_obj2dev(parent);
 791         struct dt_object *tgt_obj = NULL;
 792         struct dt_object *bak_obj = NULL;
 793         const struct dt_it_ops *iops;
 794         struct dt_it *di;
 795         loff_t pos = sizeof(struct lustre_index_backup_header);
 796         int count = 0;
 797         int size = 0;
 798         int rc;
 799         ENTRY;
 800
 801         tgt_obj = lu2dt(lu_object_find_slice(env, &dev->dd_lu_dev,
 802                                              &libu->libu_fid, NULL));
 803         if (IS_ERR_OR_NULL(tgt_obj))
 804                 GOTO(out, rc = tgt_obj ? PTR_ERR(tgt_obj) : -ENOENT);
 805
 806         if (!dt_object_exists(tgt_obj))
 807                 GOTO(out, rc = 0);
 808
 809         if (!tgt_obj->do_index_ops) {
 810                 struct dt_index_features feat;
 811
 812                 feat.dif_flags = DT_IND_UPDATE;
 813                 feat.dif_keysize_min = libu->libu_keysize;
 814                 feat.dif_keysize_max = libu->libu_keysize;
 815                 feat.dif_recsize_min = libu->libu_recsize;
 816                 feat.dif_recsize_max = libu->libu_recsize;
 817                 feat.dif_ptrsize = 4;
 818                 rc = tgt_obj->do_ops->do_index_try(env, tgt_obj, &feat);
 819                 if (rc)
 820                         GOTO(out, rc);
 821         }
 822
 823         lustre_fid2lbx(buf, &libu->libu_fid, bufsize);
 824         bak_obj = local_file_find_or_create(env, los, parent, buf,
 825                                             S_IFREG | S_IRUGO | S_IWUSR);
 826         if (IS_ERR_OR_NULL(bak_obj))
 827                 GOTO(out, rc = bak_obj ? PTR_ERR(bak_obj) : -ENOENT);
 828
 829         iops = &tgt_obj->do_index_ops->dio_it;
 830         di = iops->init(env, tgt_obj, 0);
 831         if (IS_ERR(di))
 832                 GOTO(out, rc = PTR_ERR(di));
 833
 834         rc = iops->load(env, di, 0);
 835         if (!rc)
 836                 rc = iops->next(env, di);
 837         else if (rc > 0)
 838                 rc = 0;
 839
 840         while (!rc) {
 841                 void *key;
 842                 void *rec;
 843
 844                 key = iops->key(env, di);
 845                 memcpy(&buf[size], key, libu->libu_keysize);
 846                 size += libu->libu_keysize;
 847                 rec = &buf[size];
 848                 rc = iops->rec(env, di, rec, 0);
 849                 if (rc)
 850                         GOTO(fini, rc);
 851
 852                 size += libu->libu_recsize;
 853                 count++;
 854                 if (size + libu->libu_keysize + libu->libu_recsize > bufsize) {
 855                         rc = lustre_index_backup_body(env, bak_obj, &pos,
 856                                                       buf, size);
 857                         if (rc)
 858                                 GOTO(fini, rc);
 859
 860                         size = 0;
 861                 }
 862
 863                 rc = iops->next(env, di);
 864         }
 865
 866         if (rc >= 0 && size > 0)
 867                 rc = lustre_index_backup_body(env, bak_obj, &pos, buf, size);
 868
 869         if (rc < 0)
 870                 GOTO(fini, rc);
 871
 872         rc = lustre_index_backup_header(env, bak_obj, &libu->libu_fid,
 873                                         libu->libu_keysize, libu->libu_recsize,
 874                                         buf, bufsize, count);
 875         if (!rc)
 876                 rc = lustre_index_update_lma(env, tgt_obj, buf, bufsize);
 877
 878         if (!rc && OBD_FAIL_CHECK(OBD_FAIL_OSD_INDEX_CRASH)) {
 879                 LASSERT(bufsize >= 512);
 880
 881                 pos = 0;
 882                 memset(buf, 0, 512);
 883                 lustre_index_backup_body(env, tgt_obj, &pos, buf, 512);
 884         }
 885
 886         GOTO(fini, rc);
 887
 888 fini:
 889         iops->fini(env, di);
 890 out:
 891         if (!IS_ERR_OR_NULL(tgt_obj))
 892                 dt_object_put_nocache(env, tgt_obj);
 893         if (!IS_ERR_OR_NULL(bak_obj))
 894                 dt_object_put_nocache(env, bak_obj);
 895         return rc;
 896 }
 897
 898 void lustre_index_backup(const struct lu_env *env, struct dt_device *dev,
 899                          const char *devname, struct list_head *head,
 900                          spinlock_t *lock, int *guard, bool backup)
 901 {
 902         struct lustre_index_backup_unit *libu;
 903         struct local_oid_storage *los = NULL;
 904         struct dt_object *parent = NULL;
 905         char *buf = NULL;
 906         struct lu_fid fid;
 907         int rc;
 908         ENTRY;
 909
 910         if (dev->dd_rdonly || *guard)
 911                 RETURN_EXIT;
 912
 913         spin_lock(lock);
 914         *guard = 1;
 915         spin_unlock(lock);
 916
 917         if (list_empty(head))
 918                 RETURN_EXIT;
 919
 920         /* Handle kinds of failures during mount process. */
 921         if (!dev->dd_lu_dev.ld_site || !dev->dd_lu_dev.ld_site->ls_top_dev)
 922                 backup = false;
 923
 924         if (backup) {
 925                 OBD_ALLOC_LARGE(buf, INDEX_BACKUP_BUFSIZE);
 926                 if (!buf) {
 927                         backup = false;
 928                         goto scan;
 929                 }
 930
 931                 lu_local_obj_fid(&fid, INDEX_BACKUP_OID);
 932                 parent = lu2dt(lu_object_find_slice(env, &dev->dd_lu_dev,
 933                                                     &fid, NULL));
 934                 if (IS_ERR_OR_NULL(parent)) {
 935                         CERROR("%s: failed to locate backup dir: rc = %ld\n",
 936                                devname, parent ? PTR_ERR(parent) : -ENOENT);
 937                         backup = false;
 938                         goto scan;
 939                 }
 940
 941                 lu_local_name_obj_fid(&fid, 1);
 942                 rc = local_oid_storage_init(env, dev, &fid, &los);
 943                 if (rc) {
 944                         CERROR("%s: failed to init local storage: rc = %d\n",
 945                                devname, rc);
 946                         backup = false;
 947                 }
 948         }
 949
 950 scan:
 951         spin_lock(lock);
 952         while (!list_empty(head)) {
 953                 libu = list_entry(head->next,
 954                                   struct lustre_index_backup_unit, libu_link);
 955                 list_del_init(&libu->libu_link);
 956                 spin_unlock(lock);
 957
 958                 if (backup) {
 959                         rc = lustre_index_backup_one(env, los, parent, libu,
 960                                                      buf, INDEX_BACKUP_BUFSIZE);
 961                         CDEBUG(D_WARNING, "%s: backup index "DFID": rc = %d\n",
 962                                devname, PFID(&libu->libu_fid), rc);
 963                 }
 964
 965                 OBD_FREE_PTR(libu);
 966                 spin_lock(lock);
 967         }
 968         spin_unlock(lock);
 969
 970         if (los)
 971                 local_oid_storage_fini(env, los);
 972         if (parent)
 973                 dt_object_put_nocache(env, parent);
 974         if (buf)
 975                 OBD_FREE_LARGE(buf, INDEX_BACKUP_BUFSIZE);
 976
 977         EXIT;
 978 }
 979 EXPORT_SYMBOL(lustre_index_backup);
 980
 981 int lustre_index_restore(const struct lu_env *env, struct dt_device *dev,
 982                          const struct lu_fid *parent_fid,
 983                          const struct lu_fid *tgt_fid,
 984                          const struct lu_fid *bak_fid, const char *name,
 985                          struct list_head *head, spinlock_t *lock,
 986                          char *buf, int bufsize)
 987 {
 988         struct dt_object *parent_obj = NULL;
 989         struct dt_object *tgt_obj = NULL;
 990         struct dt_object *bak_obj = NULL;
 991         struct lustre_index_backup_header *header;
 992         struct dt_index_features *feat;
 993         struct dt_object_format *dof;
 994         struct lu_attr *la;
 995         struct thandle *th;
 996         struct lu_object_conf conf;
 997         struct dt_insert_rec ent;
 998         struct lu_buf lbuf;
 999         struct lu_fid tfid;
1000         loff_t pos = 0;
1001         __u32 keysize;
1002         __u32 recsize;
1003         __u32 pairsize;
1004         int count;
1005         int rc;
1006         bool registered = false;
1007         ENTRY;
1008
1009         LASSERT(bufsize >= sizeof(*la) + sizeof(*dof) +
1010                 sizeof(*feat) + sizeof(*header));
1011
1012         memset(buf, 0, bufsize);
1013         la = (struct lu_attr *)buf;
1014         dof = (void *)la + sizeof(*la);
1015         feat = (void *)dof + sizeof(*dof);
1016         header = (void *)feat + sizeof(*feat);
1017         lbuf.lb_buf = header;
1018         lbuf.lb_len = sizeof(*header);
1019
1020         tgt_obj = lu2dt(lu_object_find_slice(env, &dev->dd_lu_dev,
1021                                              tgt_fid, NULL));
1022         if (IS_ERR_OR_NULL(tgt_obj))
1023                 GOTO(out, rc = tgt_obj ? PTR_ERR(tgt_obj) : -ENOENT);
1024
1025         bak_obj = lu2dt(lu_object_find_slice(env, &dev->dd_lu_dev,
1026                                              bak_fid, NULL));
1027         if (IS_ERR_OR_NULL(bak_obj))
1028                 GOTO(out, rc = bak_obj ? PTR_ERR(bak_obj) : -ENOENT);
1029
1030         if (!dt_object_exists(bak_obj))
1031                 GOTO(out, rc = -ENOENT);
1032
1033         parent_obj = lu2dt(lu_object_find_slice(env, &dev->dd_lu_dev,
1034                                                 parent_fid, NULL));
1035         if (IS_ERR_OR_NULL(parent_obj))
1036                 GOTO(out, rc = parent_obj ? PTR_ERR(parent_obj) : -ENOENT);
1037
1038         LASSERT(dt_object_exists(parent_obj));
1039
1040         if (unlikely(!dt_try_as_dir(env, parent_obj)))
1041                 GOTO(out, rc = -ENOTDIR);
1042
1043         rc = dt_attr_get(env, tgt_obj, la);
1044         if (rc)
1045                 GOTO(out, rc);
1046
1047         rc = dt_record_read(env, bak_obj, &lbuf, &pos);
1048         if (rc)
1049                 GOTO(out, rc);
1050
1051         if (le32_to_cpu(header->libh_magic) != INDEX_BACKUP_MAGIC_V1)
1052                 GOTO(out, rc = -EINVAL);
1053
1054         fid_le_to_cpu(&tfid, &header->libh_owner);
1055         if (unlikely(!lu_fid_eq(tgt_fid, &tfid)))
1056                 GOTO(out, rc = -EINVAL);
1057
1058         keysize = le32_to_cpu(header->libh_keysize);
1059         recsize = le32_to_cpu(header->libh_recsize);
1060         pairsize = keysize + recsize;
1061
1062         memset(feat, 0, sizeof(*feat));
1063         feat->dif_flags = DT_IND_UPDATE;
1064         feat->dif_keysize_min = feat->dif_keysize_max = keysize;
1065         feat->dif_recsize_min = feat->dif_recsize_max = recsize;
1066         feat->dif_ptrsize = 4;
1067
1068         /* T1: remove old name entry and destroy old index. */
1069         th = dt_trans_create(env, dev);
1070         if (IS_ERR(th))
1071                 GOTO(out, rc = PTR_ERR(th));
1072
1073         rc = dt_declare_delete(env, parent_obj,
1074                                (const struct dt_key *)name, th);
1075         if (rc)
1076                 GOTO(stop, rc);
1077
1078         rc = dt_declare_destroy(env, tgt_obj, th);
1079         if (rc)
1080                 GOTO(stop, rc);
1081
1082         rc = dt_trans_start_local(env, dev, th);
1083         if (rc)
1084                 GOTO(stop, rc);
1085
1086         rc = dt_delete(env, parent_obj, (const struct dt_key *)name, th);
1087         if (rc)
1088                 GOTO(stop, rc);
1089
1090         dt_write_lock(env, tgt_obj, 0);
1091         rc = dt_destroy(env, tgt_obj, th);
1092         dt_write_unlock(env, tgt_obj);
1093         dt_trans_stop(env, dev, th);
1094         if (rc)
1095                 GOTO(out, rc);
1096
1097         la->la_valid = LA_MODE | LA_UID | LA_GID;
1098         conf.loc_flags = LOC_F_NEW;
1099         dof->u.dof_idx.di_feat = feat;
1100         dof->dof_type = DFT_INDEX;
1101         ent.rec_type = S_IFREG;
1102         ent.rec_fid = tgt_fid;
1103
1104         /* Drop cache before re-create it. */
1105         dt_object_put_nocache(env, tgt_obj);
1106         tgt_obj = lu2dt(lu_object_find_slice(env, &dev->dd_lu_dev,
1107                                              tgt_fid, &conf));
1108         if (IS_ERR_OR_NULL(tgt_obj))
1109                 GOTO(out, rc = tgt_obj ? PTR_ERR(tgt_obj) : -ENOENT);
1110
1111         LASSERT(!dt_object_exists(tgt_obj));
1112
1113         /* T2: create new index and insert new name entry. */
1114         th = dt_trans_create(env, dev);
1115         if (IS_ERR(th))
1116                 GOTO(out, rc = PTR_ERR(th));
1117
1118         rc = dt_declare_create(env, tgt_obj, la, NULL, dof, th);
1119         if (rc)
1120                 GOTO(stop, rc);
1121
1122         rc = dt_declare_insert(env, parent_obj, (const struct dt_rec *)&ent,
1123                                (const struct dt_key *)name, th);
1124         if (rc)
1125                 GOTO(stop, rc);
1126
1127         rc = dt_trans_start_local(env, dev, th);
1128         if (rc)
1129                 GOTO(stop, rc);
1130
1131         dt_write_lock(env, tgt_obj, 0);
1132         rc = dt_create(env, tgt_obj, la, NULL, dof, th);
1133         dt_write_unlock(env, tgt_obj);
1134         if (rc)
1135                 GOTO(stop, rc);
1136
1137         rc = dt_insert(env, parent_obj, (const struct dt_rec *)&ent,
1138                        (const struct dt_key *)name, th);
1139         dt_trans_stop(env, dev, th);
1140         /* Some index name may has been inserted by OSD
1141          * automatically when create the index object. */
1142         if (unlikely(rc == -EEXIST))
1143                 rc = 0;
1144         if (rc)
1145                 GOTO(out, rc);
1146
1147         /* The new index will register via index_try. */
1148         rc = tgt_obj->do_ops->do_index_try(env, tgt_obj, feat);
1149         if (rc)
1150                 GOTO(out, rc);
1151
1152         registered = true;
1153         count = le32_to_cpu(header->libh_count);
1154         while (!rc && count > 0) {
1155                 int size = pairsize * count;
1156                 int items = count;
1157                 int i;
1158
1159                 if (size > bufsize) {
1160                         items = bufsize / pairsize;
1161                         size = pairsize * items;
1162                 }
1163
1164                 lbuf.lb_buf = buf;
1165                 lbuf.lb_len = size;
1166                 rc = dt_record_read(env, bak_obj, &lbuf, &pos);
1167                 for (i = 0; i < items && !rc; i++) {
1168                         void *key = &buf[i * pairsize];
1169                         void *rec = &buf[i * pairsize + keysize];
1170
1171                         /* Tn: restore the records. */
1172                         th = dt_trans_create(env, dev);
1173                         if (!th)
1174                                 GOTO(out, rc = -ENOMEM);
1175
1176                         rc = dt_declare_insert(env, tgt_obj, rec, key, th);
1177                         if (rc)
1178                                 GOTO(stop, rc);
1179
1180                         rc = dt_trans_start_local(env, dev, th);
1181                         if (rc)
1182                                 GOTO(stop, rc);
1183
1184                         rc = dt_insert(env, tgt_obj, rec, key, th);
1185                         if (unlikely(rc == -EEXIST))
1186                                 rc = 0;
1187
1188                         dt_trans_stop(env, dev, th);
1189                 }
1190
1191                 count -= items;
1192         }
1193
1194         GOTO(out, rc);
1195
1196 stop:
1197         dt_trans_stop(env, dev, th);
1198         if (rc && registered)
1199                 /* Degister the index to avoid overwriting the backup. */
1200                 lustre_index_degister(head, lock, tgt_fid);
1201
1202 out:
1203         if (!IS_ERR_OR_NULL(tgt_obj))
1204                 dt_object_put_nocache(env, tgt_obj);
1205         if (!IS_ERR_OR_NULL(bak_obj))
1206                 dt_object_put_nocache(env, bak_obj);
1207         if (!IS_ERR_OR_NULL(parent_obj))
1208                 dt_object_put_nocache(env, parent_obj);
1209         return rc;
1210 }
1211 EXPORT_SYMBOL(lustre_index_restore);