fs/nilfs2/recovery.c

   1 /*
   2  * recovery.c - NILFS recovery logic
   3  *
   4  * Copyright (C) 2005-2008 Nippon Telegraph and Telephone Corporation.
   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 as published by
   8  * the Free Software Foundation; either version 2 of the License, or
   9  * (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  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., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  19  *
  20  * Written by Ryusuke Konishi <ryusuke@osrg.net>
  21  */
  22
  23 #include <linux/buffer_head.h>
  24 #include <linux/blkdev.h>
  25 #include <linux/swap.h>
  26 #include <linux/slab.h>
  27 #include <linux/crc32.h>
  28 #include "nilfs.h"
  29 #include "segment.h"
  30 #include "sufile.h"
  31 #include "page.h"
  32 #include "segbuf.h"
  33
  34 /*
  35  * Segment check result
  36  */
  37 enum {
  38         NILFS_SEG_VALID,
  39         NILFS_SEG_NO_SUPER_ROOT,
  40         NILFS_SEG_FAIL_IO,
  41         NILFS_SEG_FAIL_MAGIC,
  42         NILFS_SEG_FAIL_SEQ,
  43         NILFS_SEG_FAIL_CHECKSUM_SUPER_ROOT,
  44         NILFS_SEG_FAIL_CHECKSUM_FULL,
  45         NILFS_SEG_FAIL_CONSISTENCY,
  46 };
  47
  48 /* work structure for recovery */
  49 struct nilfs_recovery_block {
  50         ino_t ino;              /* Inode number of the file that this block
  51                                    belongs to */
  52         sector_t blocknr;       /* block number */
  53         __u64 vblocknr;         /* virtual block number */
  54         unsigned long blkoff;   /* File offset of the data block (per block) */
  55         struct list_head list;
  56 };
  57
  58
  59 static int nilfs_warn_segment_error(int err)
  60 {
  61         switch (err) {
  62         case NILFS_SEG_FAIL_IO:
  63                 printk(KERN_WARNING
  64                        "NILFS warning: I/O error on loading last segment\n");
  65                 return -EIO;
  66         case NILFS_SEG_FAIL_MAGIC:
  67                 printk(KERN_WARNING
  68                        "NILFS warning: Segment magic number invalid\n");
  69                 break;
  70         case NILFS_SEG_FAIL_SEQ:
  71                 printk(KERN_WARNING
  72                        "NILFS warning: Sequence number mismatch\n");
  73                 break;
  74         case NILFS_SEG_FAIL_CHECKSUM_SUPER_ROOT:
  75                 printk(KERN_WARNING
  76                        "NILFS warning: Checksum error in super root\n");
  77                 break;
  78         case NILFS_SEG_FAIL_CHECKSUM_FULL:
  79                 printk(KERN_WARNING
  80                        "NILFS warning: Checksum error in segment payload\n");
  81                 break;
  82         case NILFS_SEG_FAIL_CONSISTENCY:
  83                 printk(KERN_WARNING
  84                        "NILFS warning: Inconsistent segment\n");
  85                 break;
  86         case NILFS_SEG_NO_SUPER_ROOT:
  87                 printk(KERN_WARNING
  88                        "NILFS warning: No super root in the last segment\n");
  89                 break;
  90         }
  91         return -EINVAL;
  92 }
  93
  94 static void store_segsum_info(struct nilfs_segsum_info *ssi,
  95                               struct nilfs_segment_summary *sum,
  96                               unsigned int blocksize)
  97 {
  98         ssi->flags = le16_to_cpu(sum->ss_flags);
  99         ssi->seg_seq = le64_to_cpu(sum->ss_seq);
 100         ssi->ctime = le64_to_cpu(sum->ss_create);
 101         ssi->next = le64_to_cpu(sum->ss_next);
 102         ssi->nblocks = le32_to_cpu(sum->ss_nblocks);
 103         ssi->nfinfo = le32_to_cpu(sum->ss_nfinfo);
 104         ssi->sumbytes = le32_to_cpu(sum->ss_sumbytes);
 105
 106         ssi->nsumblk = DIV_ROUND_UP(ssi->sumbytes, blocksize);
 107         ssi->nfileblk = ssi->nblocks - ssi->nsumblk - !!NILFS_SEG_HAS_SR(ssi);
 108
 109         /* need to verify ->ss_bytes field if read ->ss_cno */
 110 }
 111
 112 /**
 113  * nilfs_compute_checksum - compute checksum of blocks continuously
 114  * @nilfs: nilfs object
 115  * @bhs: buffer head of start block
 116  * @sum: place to store result
 117  * @offset: offset bytes in the first block
 118  * @check_bytes: number of bytes to be checked
 119  * @start: DBN of start block
 120  * @nblock: number of blocks to be checked
 121  */
 122 static int nilfs_compute_checksum(struct the_nilfs *nilfs,
 123                                   struct buffer_head *bhs, u32 *sum,
 124                                   unsigned long offset, u64 check_bytes,
 125                                   sector_t start, unsigned long nblock)
 126 {
 127         unsigned int blocksize = nilfs->ns_blocksize;
 128         unsigned long size;
 129         u32 crc;
 130
 131         BUG_ON(offset >= blocksize);
 132         check_bytes -= offset;
 133         size = min_t(u64, check_bytes, blocksize - offset);
 134         crc = crc32_le(nilfs->ns_crc_seed,
 135                        (unsigned char *)bhs->b_data + offset, size);
 136         if (--nblock > 0) {
 137                 do {
 138                         struct buffer_head *bh;
 139
 140                         bh = __bread(nilfs->ns_bdev, ++start, blocksize);
 141                         if (!bh)
 142                                 return -EIO;
 143                         check_bytes -= size;
 144                         size = min_t(u64, check_bytes, blocksize);
 145                         crc = crc32_le(crc, bh->b_data, size);
 146                         brelse(bh);
 147                 } while (--nblock > 0);
 148         }
 149         *sum = crc;
 150         return 0;
 151 }
 152
 153 /**
 154  * nilfs_read_super_root_block - read super root block
 155  * @nilfs: nilfs object
 156  * @sr_block: disk block number of the super root block
 157  * @pbh: address of a buffer_head pointer to return super root buffer
 158  * @check: CRC check flag
 159  */
 160 int nilfs_read_super_root_block(struct the_nilfs *nilfs, sector_t sr_block,
 161                                 struct buffer_head **pbh, int check)
 162 {
 163         struct buffer_head *bh_sr;
 164         struct nilfs_super_root *sr;
 165         u32 crc;
 166         int ret;
 167
 168         *pbh = NULL;
 169         bh_sr = __bread(nilfs->ns_bdev, sr_block, nilfs->ns_blocksize);
 170         if (unlikely(!bh_sr)) {
 171                 ret = NILFS_SEG_FAIL_IO;
 172                 goto failed;
 173         }
 174
 175         sr = (struct nilfs_super_root *)bh_sr->b_data;
 176         if (check) {
 177                 unsigned bytes = le16_to_cpu(sr->sr_bytes);
 178
 179                 if (bytes == 0 || bytes > nilfs->ns_blocksize) {
 180                         ret = NILFS_SEG_FAIL_CHECKSUM_SUPER_ROOT;
 181                         goto failed_bh;
 182                 }
 183                 if (nilfs_compute_checksum(
 184                             nilfs, bh_sr, &crc, sizeof(sr->sr_sum), bytes,
 185                             sr_block, 1)) {
 186                         ret = NILFS_SEG_FAIL_IO;
 187                         goto failed_bh;
 188                 }
 189                 if (crc != le32_to_cpu(sr->sr_sum)) {
 190                         ret = NILFS_SEG_FAIL_CHECKSUM_SUPER_ROOT;
 191                         goto failed_bh;
 192                 }
 193         }
 194         *pbh = bh_sr;
 195         return 0;
 196
 197  failed_bh:
 198         brelse(bh_sr);
 199
 200  failed:
 201         return nilfs_warn_segment_error(ret);
 202 }
 203
 204 /**
 205  * load_segment_summary - read segment summary of the specified partial segment
 206  * @nilfs: nilfs object
 207  * @pseg_start: start disk block number of partial segment
 208  * @seg_seq: sequence number requested
 209  * @ssi: pointer to nilfs_segsum_info struct to store information
 210  */
 211 static int
 212 load_segment_summary(struct the_nilfs *nilfs, sector_t pseg_start,
 213                      u64 seg_seq, struct nilfs_segsum_info *ssi)
 214 {
 215         struct buffer_head *bh_sum;
 216         struct nilfs_segment_summary *sum;
 217         unsigned long nblock;
 218         u32 crc;
 219         int ret = NILFS_SEG_FAIL_IO;
 220
 221         bh_sum = __bread(nilfs->ns_bdev, pseg_start, nilfs->ns_blocksize);
 222         if (!bh_sum)
 223                 goto out;
 224
 225         sum = (struct nilfs_segment_summary *)bh_sum->b_data;
 226
 227         /* Check consistency of segment summary */
 228         if (le32_to_cpu(sum->ss_magic) != NILFS_SEGSUM_MAGIC) {
 229                 ret = NILFS_SEG_FAIL_MAGIC;
 230                 goto failed;
 231         }
 232         store_segsum_info(ssi, sum, nilfs->ns_blocksize);
 233         if (seg_seq != ssi->seg_seq) {
 234                 ret = NILFS_SEG_FAIL_SEQ;
 235                 goto failed;
 236         }
 237
 238         nblock = ssi->nblocks;
 239         if (unlikely(nblock == 0 || nblock > nilfs->ns_blocks_per_segment)) {
 240                 /* This limits the number of blocks read in the CRC check */
 241                 ret = NILFS_SEG_FAIL_CONSISTENCY;
 242                 goto failed;
 243         }
 244         if (nilfs_compute_checksum(nilfs, bh_sum, &crc, sizeof(sum->ss_datasum),
 245                                    ((u64)nblock << nilfs->ns_blocksize_bits),
 246                                    pseg_start, nblock)) {
 247                 ret = NILFS_SEG_FAIL_IO;
 248                 goto failed;
 249         }
 250         if (crc == le32_to_cpu(sum->ss_datasum))
 251                 ret = 0;
 252         else
 253                 ret = NILFS_SEG_FAIL_CHECKSUM_FULL;
 254  failed:
 255         brelse(bh_sum);
 256  out:
 257         return ret;
 258 }
 259
 260 /**
 261  * nilfs_read_summary_info - read an item on summary blocks of a log
 262  * @nilfs: nilfs object
 263  * @pbh: the current buffer head on summary blocks [in, out]
 264  * @offset: the current byte offset on summary blocks [in, out]
 265  * @bytes: byte size of the item to be read
 266  */
 267 static void *nilfs_read_summary_info(struct the_nilfs *nilfs,
 268                                      struct buffer_head **pbh,
 269                                      unsigned int *offset, unsigned int bytes)
 270 {
 271         void *ptr;
 272         sector_t blocknr;
 273
 274         BUG_ON((*pbh)->b_size < *offset);
 275         if (bytes > (*pbh)->b_size - *offset) {
 276                 blocknr = (*pbh)->b_blocknr;
 277                 brelse(*pbh);
 278                 *pbh = __bread(nilfs->ns_bdev, blocknr + 1,
 279                                nilfs->ns_blocksize);
 280                 if (unlikely(!*pbh))
 281                         return NULL;
 282                 *offset = 0;
 283         }
 284         ptr = (*pbh)->b_data + *offset;
 285         *offset += bytes;
 286         return ptr;
 287 }
 288
 289 /**
 290  * nilfs_skip_summary_info - skip items on summary blocks of a log
 291  * @nilfs: nilfs object
 292  * @pbh: the current buffer head on summary blocks [in, out]
 293  * @offset: the current byte offset on summary blocks [in, out]
 294  * @bytes: byte size of the item to be skipped
 295  * @count: number of items to be skipped
 296  */
 297 static void nilfs_skip_summary_info(struct the_nilfs *nilfs,
 298                                     struct buffer_head **pbh,
 299                                     unsigned int *offset, unsigned int bytes,
 300                                     unsigned long count)
 301 {
 302         unsigned int rest_item_in_current_block
 303                 = ((*pbh)->b_size - *offset) / bytes;
 304
 305         if (count <= rest_item_in_current_block) {
 306                 *offset += bytes * count;
 307         } else {
 308                 sector_t blocknr = (*pbh)->b_blocknr;
 309                 unsigned int nitem_per_block = (*pbh)->b_size / bytes;
 310                 unsigned int bcnt;
 311
 312                 count -= rest_item_in_current_block;
 313                 bcnt = DIV_ROUND_UP(count, nitem_per_block);
 314                 *offset = bytes * (count - (bcnt - 1) * nitem_per_block);
 315
 316                 brelse(*pbh);
 317                 *pbh = __bread(nilfs->ns_bdev, blocknr + bcnt,
 318                                nilfs->ns_blocksize);
 319         }
 320 }
 321
 322 /**
 323  * nilfs_scan_dsync_log - get block information of a log written for data sync
 324  * @nilfs: nilfs object
 325  * @start_blocknr: start block number of the log
 326  * @ssi: log summary information
 327  * @head: list head to add nilfs_recovery_block struct
 328  */
 329 static int nilfs_scan_dsync_log(struct the_nilfs *nilfs, sector_t start_blocknr,
 330                                 struct nilfs_segsum_info *ssi,
 331                                 struct list_head *head)
 332 {
 333         struct buffer_head *bh;
 334         unsigned int offset;
 335         unsigned long nfinfo = ssi->nfinfo;
 336         sector_t blocknr = start_blocknr + ssi->nsumblk;
 337         ino_t ino;
 338         int err = -EIO;
 339
 340         if (!nfinfo)
 341                 return 0;
 342
 343         bh = __bread(nilfs->ns_bdev, start_blocknr, nilfs->ns_blocksize);
 344         if (unlikely(!bh))
 345                 goto out;
 346
 347         offset = le16_to_cpu(
 348                 ((struct nilfs_segment_summary *)bh->b_data)->ss_bytes);
 349         for (;;) {
 350                 unsigned long nblocks, ndatablk, nnodeblk;
 351                 struct nilfs_finfo *finfo;
 352
 353                 finfo = nilfs_read_summary_info(nilfs, &bh, &offset,
 354                                                 sizeof(*finfo));
 355                 if (unlikely(!finfo))
 356                         goto out;
 357
 358                 ino = le64_to_cpu(finfo->fi_ino);
 359                 nblocks = le32_to_cpu(finfo->fi_nblocks);
 360                 ndatablk = le32_to_cpu(finfo->fi_ndatablk);
 361                 nnodeblk = nblocks - ndatablk;
 362
 363                 while (ndatablk-- > 0) {
 364                         struct nilfs_recovery_block *rb;
 365                         struct nilfs_binfo_v *binfo;
 366
 367                         binfo = nilfs_read_summary_info(nilfs, &bh, &offset,
 368                                                         sizeof(*binfo));
 369                         if (unlikely(!binfo))
 370                                 goto out;
 371
 372                         rb = kmalloc(sizeof(*rb), GFP_NOFS);
 373                         if (unlikely(!rb)) {
 374                                 err = -ENOMEM;
 375                                 goto out;
 376                         }
 377                         rb->ino = ino;
 378                         rb->blocknr = blocknr++;
 379                         rb->vblocknr = le64_to_cpu(binfo->bi_vblocknr);
 380                         rb->blkoff = le64_to_cpu(binfo->bi_blkoff);
 381                         /* INIT_LIST_HEAD(&rb->list); */
 382                         list_add_tail(&rb->list, head);
 383                 }
 384                 if (--nfinfo == 0)
 385                         break;
 386                 blocknr += nnodeblk; /* always 0 for data sync logs */
 387                 nilfs_skip_summary_info(nilfs, &bh, &offset, sizeof(__le64),
 388                                         nnodeblk);
 389                 if (unlikely(!bh))
 390                         goto out;
 391         }
 392         err = 0;
 393  out:
 394         brelse(bh);   /* brelse(NULL) is just ignored */
 395         return err;
 396 }
 397
 398 static void dispose_recovery_list(struct list_head *head)
 399 {
 400         while (!list_empty(head)) {
 401                 struct nilfs_recovery_block *rb
 402                         = list_entry(head->next,
 403                                      struct nilfs_recovery_block, list);
 404                 list_del(&rb->list);
 405                 kfree(rb);
 406         }
 407 }
 408
 409 struct nilfs_segment_entry {
 410         struct list_head        list;
 411         __u64                   segnum;
 412 };
 413
 414 static int nilfs_segment_list_add(struct list_head *head, __u64 segnum)
 415 {
 416         struct nilfs_segment_entry *ent = kmalloc(sizeof(*ent), GFP_NOFS);
 417
 418         if (unlikely(!ent))
 419                 return -ENOMEM;
 420
 421         ent->segnum = segnum;
 422         INIT_LIST_HEAD(&ent->list);
 423         list_add_tail(&ent->list, head);
 424         return 0;
 425 }
 426
 427 void nilfs_dispose_segment_list(struct list_head *head)
 428 {
 429         while (!list_empty(head)) {
 430                 struct nilfs_segment_entry *ent
 431                         = list_entry(head->next,
 432                                      struct nilfs_segment_entry, list);
 433                 list_del(&ent->list);
 434                 kfree(ent);
 435         }
 436 }
 437
 438 static int nilfs_prepare_segment_for_recovery(struct the_nilfs *nilfs,
 439                                               struct nilfs_sb_info *sbi,
 440                                               struct nilfs_recovery_info *ri)
 441 {
 442         struct list_head *head = &ri->ri_used_segments;
 443         struct nilfs_segment_entry *ent, *n;
 444         struct inode *sufile = nilfs->ns_sufile;
 445         __u64 segnum[4];
 446         int err;
 447         int i;
 448
 449         segnum[0] = nilfs->ns_segnum;
 450         segnum[1] = nilfs->ns_nextnum;
 451         segnum[2] = ri->ri_segnum;
 452         segnum[3] = ri->ri_nextnum;
 453
 454         nilfs_attach_writer(nilfs, sbi);
 455         /*
 456          * Releasing the next segment of the latest super root.
 457          * The next segment is invalidated by this recovery.
 458          */
 459         err = nilfs_sufile_free(sufile, segnum[1]);
 460         if (unlikely(err))
 461                 goto failed;
 462
 463         for (i = 1; i < 4; i++) {
 464                 err = nilfs_segment_list_add(head, segnum[i]);
 465                 if (unlikely(err))
 466                         goto failed;
 467         }
 468
 469         /*
 470          * Collecting segments written after the latest super root.
 471          * These are marked dirty to avoid being reallocated in the next write.
 472          */
 473         list_for_each_entry_safe(ent, n, head, list) {
 474                 if (ent->segnum != segnum[0]) {
 475                         err = nilfs_sufile_scrap(sufile, ent->segnum);
 476                         if (unlikely(err))
 477                                 goto failed;
 478                 }
 479                 list_del(&ent->list);
 480                 kfree(ent);
 481         }
 482
 483         /* Allocate new segments for recovery */
 484         err = nilfs_sufile_alloc(sufile, &segnum[0]);
 485         if (unlikely(err))
 486                 goto failed;
 487
 488         nilfs->ns_pseg_offset = 0;
 489         nilfs->ns_seg_seq = ri->ri_seq + 2;
 490         nilfs->ns_nextnum = nilfs->ns_segnum = segnum[0];
 491
 492  failed:
 493         /* No need to recover sufile because it will be destroyed on error */
 494         nilfs_detach_writer(nilfs, sbi);
 495         return err;
 496 }
 497
 498 static int nilfs_recovery_copy_block(struct the_nilfs *nilfs,
 499                                      struct nilfs_recovery_block *rb,
 500                                      struct page *page)
 501 {
 502         struct buffer_head *bh_org;
 503         void *kaddr;
 504
 505         bh_org = __bread(nilfs->ns_bdev, rb->blocknr, nilfs->ns_blocksize);
 506         if (unlikely(!bh_org))
 507                 return -EIO;
 508
 509         kaddr = kmap_atomic(page, KM_USER0);
 510         memcpy(kaddr + bh_offset(bh_org), bh_org->b_data, bh_org->b_size);
 511         kunmap_atomic(kaddr, KM_USER0);
 512         brelse(bh_org);
 513         return 0;
 514 }
 515
 516 static int nilfs_recover_dsync_blocks(struct the_nilfs *nilfs,
 517                                       struct nilfs_sb_info *sbi,
 518                                       struct list_head *head,
 519                                       unsigned long *nr_salvaged_blocks)
 520 {
 521         struct inode *inode;
 522         struct nilfs_recovery_block *rb, *n;
 523         unsigned blocksize = nilfs->ns_blocksize;
 524         struct page *page;
 525         loff_t pos;
 526         int err = 0, err2 = 0;
 527
 528         list_for_each_entry_safe(rb, n, head, list) {
 529                 inode = nilfs_iget(sbi->s_super, rb->ino);
 530                 if (IS_ERR(inode)) {
 531                         err = PTR_ERR(inode);
 532                         inode = NULL;
 533                         goto failed_inode;
 534                 }
 535
 536                 pos = rb->blkoff << inode->i_blkbits;
 537                 page = NULL;
 538                 err = block_write_begin(NULL, inode->i_mapping, pos, blocksize,
 539                                         0, &page, NULL, nilfs_get_block);
 540                 if (unlikely(err))
 541                         goto failed_inode;
 542
 543                 err = nilfs_recovery_copy_block(nilfs, rb, page);
 544                 if (unlikely(err))
 545                         goto failed_page;
 546
 547                 err = nilfs_set_file_dirty(sbi, inode, 1);
 548                 if (unlikely(err))
 549                         goto failed_page;
 550
 551                 block_write_end(NULL, inode->i_mapping, pos, blocksize,
 552                                 blocksize, page, NULL);
 553
 554                 unlock_page(page);
 555                 page_cache_release(page);
 556
 557                 (*nr_salvaged_blocks)++;
 558                 goto next;
 559
 560  failed_page:
 561                 unlock_page(page);
 562                 page_cache_release(page);
 563
 564  failed_inode:
 565                 printk(KERN_WARNING
 566                        "NILFS warning: error recovering data block "
 567                        "(err=%d, ino=%lu, block-offset=%llu)\n",
 568                        err, (unsigned long)rb->ino,
 569                        (unsigned long long)rb->blkoff);
 570                 if (!err2)
 571                         err2 = err;
 572  next:
 573                 iput(inode); /* iput(NULL) is just ignored */
 574                 list_del_init(&rb->list);
 575                 kfree(rb);
 576         }
 577         return err2;
 578 }
 579
 580 /**
 581  * nilfs_do_roll_forward - salvage logical segments newer than the latest
 582  * checkpoint
 583  * @nilfs: nilfs object
 584  * @sbi: nilfs_sb_info
 585  * @ri: pointer to a nilfs_recovery_info
 586  */
 587 static int nilfs_do_roll_forward(struct the_nilfs *nilfs,
 588                                  struct nilfs_sb_info *sbi,
 589                                  struct nilfs_recovery_info *ri)
 590 {
 591         struct nilfs_segsum_info ssi;
 592         sector_t pseg_start;
 593         sector_t seg_start, seg_end;  /* Starting/ending DBN of full segment */
 594         unsigned long nsalvaged_blocks = 0;
 595         u64 seg_seq;
 596         __u64 segnum, nextnum = 0;
 597         int empty_seg = 0;
 598         int err = 0, ret;
 599         LIST_HEAD(dsync_blocks);  /* list of data blocks to be recovered */
 600         enum {
 601                 RF_INIT_ST,
 602                 RF_DSYNC_ST,   /* scanning data-sync segments */
 603         };
 604         int state = RF_INIT_ST;
 605
 606         nilfs_attach_writer(nilfs, sbi);
 607         pseg_start = ri->ri_lsegs_start;
 608         seg_seq = ri->ri_lsegs_start_seq;
 609         segnum = nilfs_get_segnum_of_block(nilfs, pseg_start);
 610         nilfs_get_segment_range(nilfs, segnum, &seg_start, &seg_end);
 611
 612         while (segnum != ri->ri_segnum || pseg_start <= ri->ri_pseg_start) {
 613
 614                 ret = load_segment_summary(nilfs, pseg_start, seg_seq, &ssi);
 615                 if (ret) {
 616                         if (ret == NILFS_SEG_FAIL_IO) {
 617                                 err = -EIO;
 618                                 goto failed;
 619                         }
 620                         goto strayed;
 621                 }
 622                 if (unlikely(NILFS_SEG_HAS_SR(&ssi)))
 623                         goto confused;
 624
 625                 /* Found a valid partial segment; do recovery actions */
 626                 nextnum = nilfs_get_segnum_of_block(nilfs, ssi.next);
 627                 empty_seg = 0;
 628                 nilfs->ns_ctime = ssi.ctime;
 629                 if (!(ssi.flags & NILFS_SS_GC))
 630                         nilfs->ns_nongc_ctime = ssi.ctime;
 631
 632                 switch (state) {
 633                 case RF_INIT_ST:
 634                         if (!NILFS_SEG_LOGBGN(&ssi) || !NILFS_SEG_DSYNC(&ssi))
 635                                 goto try_next_pseg;
 636                         state = RF_DSYNC_ST;
 637                         /* Fall through */
 638                 case RF_DSYNC_ST:
 639                         if (!NILFS_SEG_DSYNC(&ssi))
 640                                 goto confused;
 641
 642                         err = nilfs_scan_dsync_log(nilfs, pseg_start, &ssi,
 643                                                    &dsync_blocks);
 644                         if (unlikely(err))
 645                                 goto failed;
 646                         if (NILFS_SEG_LOGEND(&ssi)) {
 647                                 err = nilfs_recover_dsync_blocks(
 648                                         nilfs, sbi, &dsync_blocks,
 649                                         &nsalvaged_blocks);
 650                                 if (unlikely(err))
 651                                         goto failed;
 652                                 state = RF_INIT_ST;
 653                         }
 654                         break; /* Fall through to try_next_pseg */
 655                 }
 656
 657  try_next_pseg:
 658                 if (pseg_start == ri->ri_lsegs_end)
 659                         break;
 660                 pseg_start += ssi.nblocks;
 661                 if (pseg_start < seg_end)
 662                         continue;
 663                 goto feed_segment;
 664
 665  strayed:
 666                 if (pseg_start == ri->ri_lsegs_end)
 667                         break;
 668
 669  feed_segment:
 670                 /* Looking to the next full segment */
 671                 if (empty_seg++)
 672                         break;
 673                 seg_seq++;
 674                 segnum = nextnum;
 675                 nilfs_get_segment_range(nilfs, segnum, &seg_start, &seg_end);
 676                 pseg_start = seg_start;
 677         }
 678
 679         if (nsalvaged_blocks) {
 680                 printk(KERN_INFO "NILFS (device %s): salvaged %lu blocks\n",
 681                        sbi->s_super->s_id, nsalvaged_blocks);
 682                 ri->ri_need_recovery = NILFS_RECOVERY_ROLLFORWARD_DONE;
 683         }
 684  out:
 685         dispose_recovery_list(&dsync_blocks);
 686         nilfs_detach_writer(nilfs, sbi);
 687         return err;
 688
 689  confused:
 690         err = -EINVAL;
 691  failed:
 692         printk(KERN_ERR
 693                "NILFS (device %s): Error roll-forwarding "
 694                "(err=%d, pseg block=%llu). ",
 695                sbi->s_super->s_id, err, (unsigned long long)pseg_start);
 696         goto out;
 697 }
 698
 699 static void nilfs_finish_roll_forward(struct the_nilfs *nilfs,
 700                                       struct nilfs_recovery_info *ri)
 701 {
 702         struct buffer_head *bh;
 703         int err;
 704
 705         if (nilfs_get_segnum_of_block(nilfs, ri->ri_lsegs_start) !=
 706             nilfs_get_segnum_of_block(nilfs, ri->ri_super_root))
 707                 return;
 708
 709         bh = __getblk(nilfs->ns_bdev, ri->ri_lsegs_start, nilfs->ns_blocksize);
 710         BUG_ON(!bh);
 711         memset(bh->b_data, 0, bh->b_size);
 712         set_buffer_dirty(bh);
 713         err = sync_dirty_buffer(bh);
 714         if (unlikely(err))
 715                 printk(KERN_WARNING
 716                        "NILFS warning: buffer sync write failed during "
 717                        "post-cleaning of recovery.\n");
 718         brelse(bh);
 719 }
 720
 721 /**
 722  * nilfs_salvage_orphan_logs - salvage logs written after the latest checkpoint
 723  * @nilfs: nilfs object
 724  * @sbi: nilfs_sb_info
 725  * @ri: pointer to a nilfs_recovery_info struct to store search results.
 726  *
 727  * Return Value: On success, 0 is returned.  On error, one of the following
 728  * negative error code is returned.
 729  *
 730  * %-EINVAL - Inconsistent filesystem state.
 731  *
 732  * %-EIO - I/O error
 733  *
 734  * %-ENOSPC - No space left on device (only in a panic state).
 735  *
 736  * %-ERESTARTSYS - Interrupted.
 737  *
 738  * %-ENOMEM - Insufficient memory available.
 739  */
 740 int nilfs_salvage_orphan_logs(struct the_nilfs *nilfs,
 741                               struct nilfs_sb_info *sbi,
 742                               struct nilfs_recovery_info *ri)
 743 {
 744         int err;
 745
 746         if (ri->ri_lsegs_start == 0 || ri->ri_lsegs_end == 0)
 747                 return 0;
 748
 749         err = nilfs_attach_checkpoint(sbi, ri->ri_cno);
 750         if (unlikely(err)) {
 751                 printk(KERN_ERR
 752                        "NILFS: error loading the latest checkpoint.\n");
 753                 return err;
 754         }
 755
 756         err = nilfs_do_roll_forward(nilfs, sbi, ri);
 757         if (unlikely(err))
 758                 goto failed;
 759
 760         if (ri->ri_need_recovery == NILFS_RECOVERY_ROLLFORWARD_DONE) {
 761                 err = nilfs_prepare_segment_for_recovery(nilfs, sbi, ri);
 762                 if (unlikely(err)) {
 763                         printk(KERN_ERR "NILFS: Error preparing segments for "
 764                                "recovery.\n");
 765                         goto failed;
 766                 }
 767
 768                 err = nilfs_attach_segment_constructor(sbi);
 769                 if (unlikely(err))
 770                         goto failed;
 771
 772                 set_nilfs_discontinued(nilfs);
 773                 err = nilfs_construct_segment(sbi->s_super);
 774                 nilfs_detach_segment_constructor(sbi);
 775
 776                 if (unlikely(err)) {
 777                         printk(KERN_ERR "NILFS: Oops! recovery failed. "
 778                                "(err=%d)\n", err);
 779                         goto failed;
 780                 }
 781
 782                 nilfs_finish_roll_forward(nilfs, ri);
 783         }
 784
 785  failed:
 786         nilfs_detach_checkpoint(sbi);
 787         return err;
 788 }
 789
 790 /**
 791  * nilfs_search_super_root - search the latest valid super root
 792  * @nilfs: the_nilfs
 793  * @ri: pointer to a nilfs_recovery_info struct to store search results.
 794  *
 795  * nilfs_search_super_root() looks for the latest super-root from a partial
 796  * segment pointed by the superblock.  It sets up struct the_nilfs through
 797  * this search. It fills nilfs_recovery_info (ri) required for recovery.
 798  *
 799  * Return Value: On success, 0 is returned.  On error, one of the following
 800  * negative error code is returned.
 801  *
 802  * %-EINVAL - No valid segment found
 803  *
 804  * %-EIO - I/O error
 805  */
 806 int nilfs_search_super_root(struct the_nilfs *nilfs,
 807                             struct nilfs_recovery_info *ri)
 808 {
 809         struct nilfs_segsum_info ssi;
 810         sector_t pseg_start, pseg_end, sr_pseg_start = 0;
 811         sector_t seg_start, seg_end; /* range of full segment (block number) */
 812         sector_t b, end;
 813         u64 seg_seq;
 814         __u64 segnum, nextnum = 0;
 815         __u64 cno;
 816         LIST_HEAD(segments);
 817         int empty_seg = 0, scan_newer = 0;
 818         int ret;
 819
 820         pseg_start = nilfs->ns_last_pseg;
 821         seg_seq = nilfs->ns_last_seq;
 822         cno = nilfs->ns_last_cno;
 823         segnum = nilfs_get_segnum_of_block(nilfs, pseg_start);
 824
 825         /* Calculate range of segment */
 826         nilfs_get_segment_range(nilfs, segnum, &seg_start, &seg_end);
 827
 828         /* Read ahead segment */
 829         b = seg_start;
 830         while (b <= seg_end)
 831                 __breadahead(nilfs->ns_bdev, b++, nilfs->ns_blocksize);
 832
 833         for (;;) {
 834                 ret = load_segment_summary(nilfs, pseg_start, seg_seq, &ssi);
 835                 if (ret) {
 836                         if (ret == NILFS_SEG_FAIL_IO)
 837                                 goto failed;
 838                         goto strayed;
 839                 }
 840                 pseg_end = pseg_start + ssi.nblocks - 1;
 841                 if (unlikely(pseg_end > seg_end)) {
 842                         ret = NILFS_SEG_FAIL_CONSISTENCY;
 843                         goto strayed;
 844                 }
 845
 846                 /* A valid partial segment */
 847                 ri->ri_pseg_start = pseg_start;
 848                 ri->ri_seq = seg_seq;
 849                 ri->ri_segnum = segnum;
 850                 nextnum = nilfs_get_segnum_of_block(nilfs, ssi.next);
 851                 ri->ri_nextnum = nextnum;
 852                 empty_seg = 0;
 853
 854                 if (!NILFS_SEG_HAS_SR(&ssi) && !scan_newer) {
 855                         /* This will never happen because a superblock
 856                            (last_segment) always points to a pseg
 857                            having a super root. */
 858                         ret = NILFS_SEG_FAIL_CONSISTENCY;
 859                         goto failed;
 860                 }
 861
 862                 if (pseg_start == seg_start) {
 863                         nilfs_get_segment_range(nilfs, nextnum, &b, &end);
 864                         while (b <= end)
 865                                 __breadahead(nilfs->ns_bdev, b++,
 866                                              nilfs->ns_blocksize);
 867                 }
 868                 if (!NILFS_SEG_HAS_SR(&ssi)) {
 869                         if (!ri->ri_lsegs_start && NILFS_SEG_LOGBGN(&ssi)) {
 870                                 ri->ri_lsegs_start = pseg_start;
 871                                 ri->ri_lsegs_start_seq = seg_seq;
 872                         }
 873                         if (NILFS_SEG_LOGEND(&ssi))
 874                                 ri->ri_lsegs_end = pseg_start;
 875                         goto try_next_pseg;
 876                 }
 877
 878                 /* A valid super root was found. */
 879                 ri->ri_cno = cno++;
 880                 ri->ri_super_root = pseg_end;
 881                 ri->ri_lsegs_start = ri->ri_lsegs_end = 0;
 882
 883                 nilfs_dispose_segment_list(&segments);
 884                 nilfs->ns_pseg_offset = (sr_pseg_start = pseg_start)
 885                         + ssi.nblocks - seg_start;
 886                 nilfs->ns_seg_seq = seg_seq;
 887                 nilfs->ns_segnum = segnum;
 888                 nilfs->ns_cno = cno;  /* nilfs->ns_cno = ri->ri_cno + 1 */
 889                 nilfs->ns_ctime = ssi.ctime;
 890                 nilfs->ns_nextnum = nextnum;
 891
 892                 if (scan_newer)
 893                         ri->ri_need_recovery = NILFS_RECOVERY_SR_UPDATED;
 894                 else {
 895                         if (nilfs->ns_mount_state & NILFS_VALID_FS)
 896                                 goto super_root_found;
 897                         scan_newer = 1;
 898                 }
 899
 900                 /* reset region for roll-forward */
 901                 pseg_start += ssi.nblocks;
 902                 if (pseg_start < seg_end)
 903                         continue;
 904                 goto feed_segment;
 905
 906  try_next_pseg:
 907                 /* Standing on a course, or met an inconsistent state */
 908                 pseg_start += ssi.nblocks;
 909                 if (pseg_start < seg_end)
 910                         continue;
 911                 goto feed_segment;
 912
 913  strayed:
 914                 /* Off the trail */
 915                 if (!scan_newer)
 916                         /*
 917                          * This can happen if a checkpoint was written without
 918                          * barriers, or as a result of an I/O failure.
 919                          */
 920                         goto failed;
 921
 922  feed_segment:
 923                 /* Looking to the next full segment */
 924                 if (empty_seg++)
 925                         goto super_root_found; /* found a valid super root */
 926
 927                 ret = nilfs_segment_list_add(&segments, segnum);
 928                 if (unlikely(ret))
 929                         goto failed;
 930
 931                 seg_seq++;
 932                 segnum = nextnum;
 933                 nilfs_get_segment_range(nilfs, segnum, &seg_start, &seg_end);
 934                 pseg_start = seg_start;
 935         }
 936
 937  super_root_found:
 938         /* Updating pointers relating to the latest checkpoint */
 939         list_splice_tail(&segments, &ri->ri_used_segments);
 940         nilfs->ns_last_pseg = sr_pseg_start;
 941         nilfs->ns_last_seq = nilfs->ns_seg_seq;
 942         nilfs->ns_last_cno = ri->ri_cno;
 943         return 0;
 944
 945  failed:
 946         nilfs_dispose_segment_list(&segments);
 947         return (ret < 0) ? ret : nilfs_warn_segment_error(ret);
 948 }