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  * Written by Ryusuke Konishi <ryusuke@osrg.net>
  17  */
  18
  19 #include <linux/buffer_head.h>
  20 #include <linux/blkdev.h>
  21 #include <linux/swap.h>
  22 #include <linux/slab.h>
  23 #include <linux/crc32.h>
  24 #include "nilfs.h"
  25 #include "segment.h"
  26 #include "sufile.h"
  27 #include "page.h"
  28 #include "segbuf.h"
  29
  30 /*
  31  * Segment check result
  32  */
  33 enum {
  34         NILFS_SEG_VALID,
  35         NILFS_SEG_NO_SUPER_ROOT,
  36         NILFS_SEG_FAIL_IO,
  37         NILFS_SEG_FAIL_MAGIC,
  38         NILFS_SEG_FAIL_SEQ,
  39         NILFS_SEG_FAIL_CHECKSUM_SUPER_ROOT,
  40         NILFS_SEG_FAIL_CHECKSUM_FULL,
  41         NILFS_SEG_FAIL_CONSISTENCY,
  42 };
  43
  44 /* work structure for recovery */
  45 struct nilfs_recovery_block {
  46         ino_t ino;              /* Inode number of the file that this block
  47                                    belongs to */
  48         sector_t blocknr;       /* block number */
  49         __u64 vblocknr;         /* virtual block number */
  50         unsigned long blkoff;   /* File offset of the data block (per block) */
  51         struct list_head list;
  52 };
  53
  54
  55 static int nilfs_warn_segment_error(int err)
  56 {
  57         switch (err) {
  58         case NILFS_SEG_FAIL_IO:
  59                 printk(KERN_WARNING
  60                        "NILFS warning: I/O error on loading last segment\n");
  61                 return -EIO;
  62         case NILFS_SEG_FAIL_MAGIC:
  63                 printk(KERN_WARNING
  64                        "NILFS warning: Segment magic number invalid\n");
  65                 break;
  66         case NILFS_SEG_FAIL_SEQ:
  67                 printk(KERN_WARNING
  68                        "NILFS warning: Sequence number mismatch\n");
  69                 break;
  70         case NILFS_SEG_FAIL_CHECKSUM_SUPER_ROOT:
  71                 printk(KERN_WARNING
  72                        "NILFS warning: Checksum error in super root\n");
  73                 break;
  74         case NILFS_SEG_FAIL_CHECKSUM_FULL:
  75                 printk(KERN_WARNING
  76                        "NILFS warning: Checksum error in segment payload\n");
  77                 break;
  78         case NILFS_SEG_FAIL_CONSISTENCY:
  79                 printk(KERN_WARNING
  80                        "NILFS warning: Inconsistent segment\n");
  81                 break;
  82         case NILFS_SEG_NO_SUPER_ROOT:
  83                 printk(KERN_WARNING
  84                        "NILFS warning: No super root in the last segment\n");
  85                 break;
  86         }
  87         return -EINVAL;
  88 }
  89
  90 /**
  91  * nilfs_compute_checksum - compute checksum of blocks continuously
  92  * @nilfs: nilfs object
  93  * @bhs: buffer head of start block
  94  * @sum: place to store result
  95  * @offset: offset bytes in the first block
  96  * @check_bytes: number of bytes to be checked
  97  * @start: DBN of start block
  98  * @nblock: number of blocks to be checked
  99  */
 100 static int nilfs_compute_checksum(struct the_nilfs *nilfs,
 101                                   struct buffer_head *bhs, u32 *sum,
 102                                   unsigned long offset, u64 check_bytes,
 103                                   sector_t start, unsigned long nblock)
 104 {
 105         unsigned int blocksize = nilfs->ns_blocksize;
 106         unsigned long size;
 107         u32 crc;
 108
 109         BUG_ON(offset >= blocksize);
 110         check_bytes -= offset;
 111         size = min_t(u64, check_bytes, blocksize - offset);
 112         crc = crc32_le(nilfs->ns_crc_seed,
 113                        (unsigned char *)bhs->b_data + offset, size);
 114         if (--nblock > 0) {
 115                 do {
 116                         struct buffer_head *bh;
 117
 118                         bh = __bread(nilfs->ns_bdev, ++start, blocksize);
 119                         if (!bh)
 120                                 return -EIO;
 121                         check_bytes -= size;
 122                         size = min_t(u64, check_bytes, blocksize);
 123                         crc = crc32_le(crc, bh->b_data, size);
 124                         brelse(bh);
 125                 } while (--nblock > 0);
 126         }
 127         *sum = crc;
 128         return 0;
 129 }
 130
 131 /**
 132  * nilfs_read_super_root_block - read super root block
 133  * @nilfs: nilfs object
 134  * @sr_block: disk block number of the super root block
 135  * @pbh: address of a buffer_head pointer to return super root buffer
 136  * @check: CRC check flag
 137  */
 138 int nilfs_read_super_root_block(struct the_nilfs *nilfs, sector_t sr_block,
 139                                 struct buffer_head **pbh, int check)
 140 {
 141         struct buffer_head *bh_sr;
 142         struct nilfs_super_root *sr;
 143         u32 crc;
 144         int ret;
 145
 146         *pbh = NULL;
 147         bh_sr = __bread(nilfs->ns_bdev, sr_block, nilfs->ns_blocksize);
 148         if (unlikely(!bh_sr)) {
 149                 ret = NILFS_SEG_FAIL_IO;
 150                 goto failed;
 151         }
 152
 153         sr = (struct nilfs_super_root *)bh_sr->b_data;
 154         if (check) {
 155                 unsigned bytes = le16_to_cpu(sr->sr_bytes);
 156
 157                 if (bytes == 0 || bytes > nilfs->ns_blocksize) {
 158                         ret = NILFS_SEG_FAIL_CHECKSUM_SUPER_ROOT;
 159                         goto failed_bh;
 160                 }
 161                 if (nilfs_compute_checksum(
 162                             nilfs, bh_sr, &crc, sizeof(sr->sr_sum), bytes,
 163                             sr_block, 1)) {
 164                         ret = NILFS_SEG_FAIL_IO;
 165                         goto failed_bh;
 166                 }
 167                 if (crc != le32_to_cpu(sr->sr_sum)) {
 168                         ret = NILFS_SEG_FAIL_CHECKSUM_SUPER_ROOT;
 169                         goto failed_bh;
 170                 }
 171         }
 172         *pbh = bh_sr;
 173         return 0;
 174
 175  failed_bh:
 176         brelse(bh_sr);
 177
 178  failed:
 179         return nilfs_warn_segment_error(ret);
 180 }
 181
 182 /**
 183  * nilfs_read_log_header - read summary header of the specified log
 184  * @nilfs: nilfs object
 185  * @start_blocknr: start block number of the log
 186  * @sum: pointer to return segment summary structure
 187  */
 188 static struct buffer_head *
 189 nilfs_read_log_header(struct the_nilfs *nilfs, sector_t start_blocknr,
 190                       struct nilfs_segment_summary **sum)
 191 {
 192         struct buffer_head *bh_sum;
 193
 194         bh_sum = __bread(nilfs->ns_bdev, start_blocknr, nilfs->ns_blocksize);
 195         if (bh_sum)
 196                 *sum = (struct nilfs_segment_summary *)bh_sum->b_data;
 197         return bh_sum;
 198 }
 199
 200 /**
 201  * nilfs_validate_log - verify consistency of log
 202  * @nilfs: nilfs object
 203  * @seg_seq: sequence number of segment
 204  * @bh_sum: buffer head of summary block
 205  * @sum: segment summary struct
 206  */
 207 static int nilfs_validate_log(struct the_nilfs *nilfs, u64 seg_seq,
 208                               struct buffer_head *bh_sum,
 209                               struct nilfs_segment_summary *sum)
 210 {
 211         unsigned long nblock;
 212         u32 crc;
 213         int ret;
 214
 215         ret = NILFS_SEG_FAIL_MAGIC;
 216         if (le32_to_cpu(sum->ss_magic) != NILFS_SEGSUM_MAGIC)
 217                 goto out;
 218
 219         ret = NILFS_SEG_FAIL_SEQ;
 220         if (le64_to_cpu(sum->ss_seq) != seg_seq)
 221                 goto out;
 222
 223         nblock = le32_to_cpu(sum->ss_nblocks);
 224         ret = NILFS_SEG_FAIL_CONSISTENCY;
 225         if (unlikely(nblock == 0 || nblock > nilfs->ns_blocks_per_segment))
 226                 /* This limits the number of blocks read in the CRC check */
 227                 goto out;
 228
 229         ret = NILFS_SEG_FAIL_IO;
 230         if (nilfs_compute_checksum(nilfs, bh_sum, &crc, sizeof(sum->ss_datasum),
 231                                    ((u64)nblock << nilfs->ns_blocksize_bits),
 232                                    bh_sum->b_blocknr, nblock))
 233                 goto out;
 234
 235         ret = NILFS_SEG_FAIL_CHECKSUM_FULL;
 236         if (crc != le32_to_cpu(sum->ss_datasum))
 237                 goto out;
 238         ret = 0;
 239 out:
 240         return ret;
 241 }
 242
 243 /**
 244  * nilfs_read_summary_info - read an item on summary blocks of a log
 245  * @nilfs: nilfs object
 246  * @pbh: the current buffer head on summary blocks [in, out]
 247  * @offset: the current byte offset on summary blocks [in, out]
 248  * @bytes: byte size of the item to be read
 249  */
 250 static void *nilfs_read_summary_info(struct the_nilfs *nilfs,
 251                                      struct buffer_head **pbh,
 252                                      unsigned int *offset, unsigned int bytes)
 253 {
 254         void *ptr;
 255         sector_t blocknr;
 256
 257         BUG_ON((*pbh)->b_size < *offset);
 258         if (bytes > (*pbh)->b_size - *offset) {
 259                 blocknr = (*pbh)->b_blocknr;
 260                 brelse(*pbh);
 261                 *pbh = __bread(nilfs->ns_bdev, blocknr + 1,
 262                                nilfs->ns_blocksize);
 263                 if (unlikely(!*pbh))
 264                         return NULL;
 265                 *offset = 0;
 266         }
 267         ptr = (*pbh)->b_data + *offset;
 268         *offset += bytes;
 269         return ptr;
 270 }
 271
 272 /**
 273  * nilfs_skip_summary_info - skip items on summary blocks of a log
 274  * @nilfs: nilfs object
 275  * @pbh: the current buffer head on summary blocks [in, out]
 276  * @offset: the current byte offset on summary blocks [in, out]
 277  * @bytes: byte size of the item to be skipped
 278  * @count: number of items to be skipped
 279  */
 280 static void nilfs_skip_summary_info(struct the_nilfs *nilfs,
 281                                     struct buffer_head **pbh,
 282                                     unsigned int *offset, unsigned int bytes,
 283                                     unsigned long count)
 284 {
 285         unsigned int rest_item_in_current_block
 286                 = ((*pbh)->b_size - *offset) / bytes;
 287
 288         if (count <= rest_item_in_current_block) {
 289                 *offset += bytes * count;
 290         } else {
 291                 sector_t blocknr = (*pbh)->b_blocknr;
 292                 unsigned int nitem_per_block = (*pbh)->b_size / bytes;
 293                 unsigned int bcnt;
 294
 295                 count -= rest_item_in_current_block;
 296                 bcnt = DIV_ROUND_UP(count, nitem_per_block);
 297                 *offset = bytes * (count - (bcnt - 1) * nitem_per_block);
 298
 299                 brelse(*pbh);
 300                 *pbh = __bread(nilfs->ns_bdev, blocknr + bcnt,
 301                                nilfs->ns_blocksize);
 302         }
 303 }
 304
 305 /**
 306  * nilfs_scan_dsync_log - get block information of a log written for data sync
 307  * @nilfs: nilfs object
 308  * @start_blocknr: start block number of the log
 309  * @sum: log summary information
 310  * @head: list head to add nilfs_recovery_block struct
 311  */
 312 static int nilfs_scan_dsync_log(struct the_nilfs *nilfs, sector_t start_blocknr,
 313                                 struct nilfs_segment_summary *sum,
 314                                 struct list_head *head)
 315 {
 316         struct buffer_head *bh;
 317         unsigned int offset;
 318         u32 nfinfo, sumbytes;
 319         sector_t blocknr;
 320         ino_t ino;
 321         int err = -EIO;
 322
 323         nfinfo = le32_to_cpu(sum->ss_nfinfo);
 324         if (!nfinfo)
 325                 return 0;
 326
 327         sumbytes = le32_to_cpu(sum->ss_sumbytes);
 328         blocknr = start_blocknr + DIV_ROUND_UP(sumbytes, nilfs->ns_blocksize);
 329         bh = __bread(nilfs->ns_bdev, start_blocknr, nilfs->ns_blocksize);
 330         if (unlikely(!bh))
 331                 goto out;
 332
 333         offset = le16_to_cpu(sum->ss_bytes);
 334         for (;;) {
 335                 unsigned long nblocks, ndatablk, nnodeblk;
 336                 struct nilfs_finfo *finfo;
 337
 338                 finfo = nilfs_read_summary_info(nilfs, &bh, &offset,
 339                                                 sizeof(*finfo));
 340                 if (unlikely(!finfo))
 341                         goto out;
 342
 343                 ino = le64_to_cpu(finfo->fi_ino);
 344                 nblocks = le32_to_cpu(finfo->fi_nblocks);
 345                 ndatablk = le32_to_cpu(finfo->fi_ndatablk);
 346                 nnodeblk = nblocks - ndatablk;
 347
 348                 while (ndatablk-- > 0) {
 349                         struct nilfs_recovery_block *rb;
 350                         struct nilfs_binfo_v *binfo;
 351
 352                         binfo = nilfs_read_summary_info(nilfs, &bh, &offset,
 353                                                         sizeof(*binfo));
 354                         if (unlikely(!binfo))
 355                                 goto out;
 356
 357                         rb = kmalloc(sizeof(*rb), GFP_NOFS);
 358                         if (unlikely(!rb)) {
 359                                 err = -ENOMEM;
 360                                 goto out;
 361                         }
 362                         rb->ino = ino;
 363                         rb->blocknr = blocknr++;
 364                         rb->vblocknr = le64_to_cpu(binfo->bi_vblocknr);
 365                         rb->blkoff = le64_to_cpu(binfo->bi_blkoff);
 366                         /* INIT_LIST_HEAD(&rb->list); */
 367                         list_add_tail(&rb->list, head);
 368                 }
 369                 if (--nfinfo == 0)
 370                         break;
 371                 blocknr += nnodeblk; /* always 0 for data sync logs */
 372                 nilfs_skip_summary_info(nilfs, &bh, &offset, sizeof(__le64),
 373                                         nnodeblk);
 374                 if (unlikely(!bh))
 375                         goto out;
 376         }
 377         err = 0;
 378  out:
 379         brelse(bh);   /* brelse(NULL) is just ignored */
 380         return err;
 381 }
 382
 383 static void dispose_recovery_list(struct list_head *head)
 384 {
 385         while (!list_empty(head)) {
 386                 struct nilfs_recovery_block *rb;
 387
 388                 rb = list_first_entry(head, struct nilfs_recovery_block, list);
 389                 list_del(&rb->list);
 390                 kfree(rb);
 391         }
 392 }
 393
 394 struct nilfs_segment_entry {
 395         struct list_head        list;
 396         __u64                   segnum;
 397 };
 398
 399 static int nilfs_segment_list_add(struct list_head *head, __u64 segnum)
 400 {
 401         struct nilfs_segment_entry *ent = kmalloc(sizeof(*ent), GFP_NOFS);
 402
 403         if (unlikely(!ent))
 404                 return -ENOMEM;
 405
 406         ent->segnum = segnum;
 407         INIT_LIST_HEAD(&ent->list);
 408         list_add_tail(&ent->list, head);
 409         return 0;
 410 }
 411
 412 void nilfs_dispose_segment_list(struct list_head *head)
 413 {
 414         while (!list_empty(head)) {
 415                 struct nilfs_segment_entry *ent;
 416
 417                 ent = list_first_entry(head, struct nilfs_segment_entry, list);
 418                 list_del(&ent->list);
 419                 kfree(ent);
 420         }
 421 }
 422
 423 static int nilfs_prepare_segment_for_recovery(struct the_nilfs *nilfs,
 424                                               struct super_block *sb,
 425                                               struct nilfs_recovery_info *ri)
 426 {
 427         struct list_head *head = &ri->ri_used_segments;
 428         struct nilfs_segment_entry *ent, *n;
 429         struct inode *sufile = nilfs->ns_sufile;
 430         __u64 segnum[4];
 431         int err;
 432         int i;
 433
 434         segnum[0] = nilfs->ns_segnum;
 435         segnum[1] = nilfs->ns_nextnum;
 436         segnum[2] = ri->ri_segnum;
 437         segnum[3] = ri->ri_nextnum;
 438
 439         /*
 440          * Releasing the next segment of the latest super root.
 441          * The next segment is invalidated by this recovery.
 442          */
 443         err = nilfs_sufile_free(sufile, segnum[1]);
 444         if (unlikely(err))
 445                 goto failed;
 446
 447         for (i = 1; i < 4; i++) {
 448                 err = nilfs_segment_list_add(head, segnum[i]);
 449                 if (unlikely(err))
 450                         goto failed;
 451         }
 452
 453         /*
 454          * Collecting segments written after the latest super root.
 455          * These are marked dirty to avoid being reallocated in the next write.
 456          */
 457         list_for_each_entry_safe(ent, n, head, list) {
 458                 if (ent->segnum != segnum[0]) {
 459                         err = nilfs_sufile_scrap(sufile, ent->segnum);
 460                         if (unlikely(err))
 461                                 goto failed;
 462                 }
 463                 list_del(&ent->list);
 464                 kfree(ent);
 465         }
 466
 467         /* Allocate new segments for recovery */
 468         err = nilfs_sufile_alloc(sufile, &segnum[0]);
 469         if (unlikely(err))
 470                 goto failed;
 471
 472         nilfs->ns_pseg_offset = 0;
 473         nilfs->ns_seg_seq = ri->ri_seq + 2;
 474         nilfs->ns_nextnum = nilfs->ns_segnum = segnum[0];
 475
 476  failed:
 477         /* No need to recover sufile because it will be destroyed on error */
 478         return err;
 479 }
 480
 481 static int nilfs_recovery_copy_block(struct the_nilfs *nilfs,
 482                                      struct nilfs_recovery_block *rb,
 483                                      struct page *page)
 484 {
 485         struct buffer_head *bh_org;
 486         void *kaddr;
 487
 488         bh_org = __bread(nilfs->ns_bdev, rb->blocknr, nilfs->ns_blocksize);
 489         if (unlikely(!bh_org))
 490                 return -EIO;
 491
 492         kaddr = kmap_atomic(page);
 493         memcpy(kaddr + bh_offset(bh_org), bh_org->b_data, bh_org->b_size);
 494         kunmap_atomic(kaddr);
 495         brelse(bh_org);
 496         return 0;
 497 }
 498
 499 static int nilfs_recover_dsync_blocks(struct the_nilfs *nilfs,
 500                                       struct super_block *sb,
 501                                       struct nilfs_root *root,
 502                                       struct list_head *head,
 503                                       unsigned long *nr_salvaged_blocks)
 504 {
 505         struct inode *inode;
 506         struct nilfs_recovery_block *rb, *n;
 507         unsigned blocksize = nilfs->ns_blocksize;
 508         struct page *page;
 509         loff_t pos;
 510         int err = 0, err2 = 0;
 511
 512         list_for_each_entry_safe(rb, n, head, list) {
 513                 inode = nilfs_iget(sb, root, rb->ino);
 514                 if (IS_ERR(inode)) {
 515                         err = PTR_ERR(inode);
 516                         inode = NULL;
 517                         goto failed_inode;
 518                 }
 519
 520                 pos = rb->blkoff << inode->i_blkbits;
 521                 err = block_write_begin(inode->i_mapping, pos, blocksize,
 522                                         0, &page, nilfs_get_block);
 523                 if (unlikely(err)) {
 524                         loff_t isize = inode->i_size;
 525                         if (pos + blocksize > isize)
 526                                 nilfs_write_failed(inode->i_mapping,
 527                                                         pos + blocksize);
 528                         goto failed_inode;
 529                 }
 530
 531                 err = nilfs_recovery_copy_block(nilfs, rb, page);
 532                 if (unlikely(err))
 533                         goto failed_page;
 534
 535                 err = nilfs_set_file_dirty(inode, 1);
 536                 if (unlikely(err))
 537                         goto failed_page;
 538
 539                 block_write_end(NULL, inode->i_mapping, pos, blocksize,
 540                                 blocksize, page, NULL);
 541
 542                 unlock_page(page);
 543                 put_page(page);
 544
 545                 (*nr_salvaged_blocks)++;
 546                 goto next;
 547
 548  failed_page:
 549                 unlock_page(page);
 550                 put_page(page);
 551
 552  failed_inode:
 553                 printk(KERN_WARNING
 554                        "NILFS warning: error recovering data block "
 555                        "(err=%d, ino=%lu, block-offset=%llu)\n",
 556                        err, (unsigned long)rb->ino,
 557                        (unsigned long long)rb->blkoff);
 558                 if (!err2)
 559                         err2 = err;
 560  next:
 561                 iput(inode); /* iput(NULL) is just ignored */
 562                 list_del_init(&rb->list);
 563                 kfree(rb);
 564         }
 565         return err2;
 566 }
 567
 568 /**
 569  * nilfs_do_roll_forward - salvage logical segments newer than the latest
 570  * checkpoint
 571  * @nilfs: nilfs object
 572  * @sb: super block instance
 573  * @ri: pointer to a nilfs_recovery_info
 574  */
 575 static int nilfs_do_roll_forward(struct the_nilfs *nilfs,
 576                                  struct super_block *sb,
 577                                  struct nilfs_root *root,
 578                                  struct nilfs_recovery_info *ri)
 579 {
 580         struct buffer_head *bh_sum = NULL;
 581         struct nilfs_segment_summary *sum = NULL;
 582         sector_t pseg_start;
 583         sector_t seg_start, seg_end;  /* Starting/ending DBN of full segment */
 584         unsigned long nsalvaged_blocks = 0;
 585         unsigned int flags;
 586         u64 seg_seq;
 587         __u64 segnum, nextnum = 0;
 588         int empty_seg = 0;
 589         int err = 0, ret;
 590         LIST_HEAD(dsync_blocks);  /* list of data blocks to be recovered */
 591         enum {
 592                 RF_INIT_ST,
 593                 RF_DSYNC_ST,   /* scanning data-sync segments */
 594         };
 595         int state = RF_INIT_ST;
 596
 597         pseg_start = ri->ri_lsegs_start;
 598         seg_seq = ri->ri_lsegs_start_seq;
 599         segnum = nilfs_get_segnum_of_block(nilfs, pseg_start);
 600         nilfs_get_segment_range(nilfs, segnum, &seg_start, &seg_end);
 601
 602         while (segnum != ri->ri_segnum || pseg_start <= ri->ri_pseg_start) {
 603                 brelse(bh_sum);
 604                 bh_sum = nilfs_read_log_header(nilfs, pseg_start, &sum);
 605                 if (!bh_sum) {
 606                         err = -EIO;
 607                         goto failed;
 608                 }
 609
 610                 ret = nilfs_validate_log(nilfs, seg_seq, bh_sum, sum);
 611                 if (ret) {
 612                         if (ret == NILFS_SEG_FAIL_IO) {
 613                                 err = -EIO;
 614                                 goto failed;
 615                         }
 616                         goto strayed;
 617                 }
 618
 619                 flags = le16_to_cpu(sum->ss_flags);
 620                 if (flags & NILFS_SS_SR)
 621                         goto confused;
 622
 623                 /* Found a valid partial segment; do recovery actions */
 624                 nextnum = nilfs_get_segnum_of_block(nilfs,
 625                                                     le64_to_cpu(sum->ss_next));
 626                 empty_seg = 0;
 627                 nilfs->ns_ctime = le64_to_cpu(sum->ss_create);
 628                 if (!(flags & NILFS_SS_GC))
 629                         nilfs->ns_nongc_ctime = nilfs->ns_ctime;
 630
 631                 switch (state) {
 632                 case RF_INIT_ST:
 633                         if (!(flags & NILFS_SS_LOGBGN) ||
 634                             !(flags & NILFS_SS_SYNDT))
 635                                 goto try_next_pseg;
 636                         state = RF_DSYNC_ST;
 637                         /* Fall through */
 638                 case RF_DSYNC_ST:
 639                         if (!(flags & NILFS_SS_SYNDT))
 640                                 goto confused;
 641
 642                         err = nilfs_scan_dsync_log(nilfs, pseg_start, sum,
 643                                                    &dsync_blocks);
 644                         if (unlikely(err))
 645                                 goto failed;
 646                         if (flags & NILFS_SS_LOGEND) {
 647                                 err = nilfs_recover_dsync_blocks(
 648                                         nilfs, sb, root, &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 += le32_to_cpu(sum->ss_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                        sb->s_id, nsalvaged_blocks);
 682                 ri->ri_need_recovery = NILFS_RECOVERY_ROLLFORWARD_DONE;
 683         }
 684  out:
 685         brelse(bh_sum);
 686         dispose_recovery_list(&dsync_blocks);
 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                sb->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  * @sb: super block instance
 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 super_block *sb,
 742                               struct nilfs_recovery_info *ri)
 743 {
 744         struct nilfs_root *root;
 745         int err;
 746
 747         if (ri->ri_lsegs_start == 0 || ri->ri_lsegs_end == 0)
 748                 return 0;
 749
 750         err = nilfs_attach_checkpoint(sb, ri->ri_cno, true, &root);
 751         if (unlikely(err)) {
 752                 printk(KERN_ERR
 753                        "NILFS: error loading the latest checkpoint.\n");
 754                 return err;
 755         }
 756
 757         err = nilfs_do_roll_forward(nilfs, sb, root, ri);
 758         if (unlikely(err))
 759                 goto failed;
 760
 761         if (ri->ri_need_recovery == NILFS_RECOVERY_ROLLFORWARD_DONE) {
 762                 err = nilfs_prepare_segment_for_recovery(nilfs, sb, ri);
 763                 if (unlikely(err)) {
 764                         printk(KERN_ERR "NILFS: Error preparing segments for "
 765                                "recovery.\n");
 766                         goto failed;
 767                 }
 768
 769                 err = nilfs_attach_log_writer(sb, root);
 770                 if (unlikely(err))
 771                         goto failed;
 772
 773                 set_nilfs_discontinued(nilfs);
 774                 err = nilfs_construct_segment(sb);
 775                 nilfs_detach_log_writer(sb);
 776
 777                 if (unlikely(err)) {
 778                         printk(KERN_ERR "NILFS: Oops! recovery failed. "
 779                                "(err=%d)\n", err);
 780                         goto failed;
 781                 }
 782
 783                 nilfs_finish_roll_forward(nilfs, ri);
 784         }
 785
 786  failed:
 787         nilfs_put_root(root);
 788         return err;
 789 }
 790
 791 /**
 792  * nilfs_search_super_root - search the latest valid super root
 793  * @nilfs: the_nilfs
 794  * @ri: pointer to a nilfs_recovery_info struct to store search results.
 795  *
 796  * nilfs_search_super_root() looks for the latest super-root from a partial
 797  * segment pointed by the superblock.  It sets up struct the_nilfs through
 798  * this search. It fills nilfs_recovery_info (ri) required for recovery.
 799  *
 800  * Return Value: On success, 0 is returned.  On error, one of the following
 801  * negative error code is returned.
 802  *
 803  * %-EINVAL - No valid segment found
 804  *
 805  * %-EIO - I/O error
 806  *
 807  * %-ENOMEM - Insufficient memory available.
 808  */
 809 int nilfs_search_super_root(struct the_nilfs *nilfs,
 810                             struct nilfs_recovery_info *ri)
 811 {
 812         struct buffer_head *bh_sum = NULL;
 813         struct nilfs_segment_summary *sum = NULL;
 814         sector_t pseg_start, pseg_end, sr_pseg_start = 0;
 815         sector_t seg_start, seg_end; /* range of full segment (block number) */
 816         sector_t b, end;
 817         unsigned long nblocks;
 818         unsigned int flags;
 819         u64 seg_seq;
 820         __u64 segnum, nextnum = 0;
 821         __u64 cno;
 822         LIST_HEAD(segments);
 823         int empty_seg = 0, scan_newer = 0;
 824         int ret;
 825
 826         pseg_start = nilfs->ns_last_pseg;
 827         seg_seq = nilfs->ns_last_seq;
 828         cno = nilfs->ns_last_cno;
 829         segnum = nilfs_get_segnum_of_block(nilfs, pseg_start);
 830
 831         /* Calculate range of segment */
 832         nilfs_get_segment_range(nilfs, segnum, &seg_start, &seg_end);
 833
 834         /* Read ahead segment */
 835         b = seg_start;
 836         while (b <= seg_end)
 837                 __breadahead(nilfs->ns_bdev, b++, nilfs->ns_blocksize);
 838
 839         for (;;) {
 840                 brelse(bh_sum);
 841                 ret = NILFS_SEG_FAIL_IO;
 842                 bh_sum = nilfs_read_log_header(nilfs, pseg_start, &sum);
 843                 if (!bh_sum)
 844                         goto failed;
 845
 846                 ret = nilfs_validate_log(nilfs, seg_seq, bh_sum, sum);
 847                 if (ret) {
 848                         if (ret == NILFS_SEG_FAIL_IO)
 849                                 goto failed;
 850                         goto strayed;
 851                 }
 852
 853                 nblocks = le32_to_cpu(sum->ss_nblocks);
 854                 pseg_end = pseg_start + nblocks - 1;
 855                 if (unlikely(pseg_end > seg_end)) {
 856                         ret = NILFS_SEG_FAIL_CONSISTENCY;
 857                         goto strayed;
 858                 }
 859
 860                 /* A valid partial segment */
 861                 ri->ri_pseg_start = pseg_start;
 862                 ri->ri_seq = seg_seq;
 863                 ri->ri_segnum = segnum;
 864                 nextnum = nilfs_get_segnum_of_block(nilfs,
 865                                                     le64_to_cpu(sum->ss_next));
 866                 ri->ri_nextnum = nextnum;
 867                 empty_seg = 0;
 868
 869                 flags = le16_to_cpu(sum->ss_flags);
 870                 if (!(flags & NILFS_SS_SR) && !scan_newer) {
 871                         /* This will never happen because a superblock
 872                            (last_segment) always points to a pseg
 873                            having a super root. */
 874                         ret = NILFS_SEG_FAIL_CONSISTENCY;
 875                         goto failed;
 876                 }
 877
 878                 if (pseg_start == seg_start) {
 879                         nilfs_get_segment_range(nilfs, nextnum, &b, &end);
 880                         while (b <= end)
 881                                 __breadahead(nilfs->ns_bdev, b++,
 882                                              nilfs->ns_blocksize);
 883                 }
 884                 if (!(flags & NILFS_SS_SR)) {
 885                         if (!ri->ri_lsegs_start && (flags & NILFS_SS_LOGBGN)) {
 886                                 ri->ri_lsegs_start = pseg_start;
 887                                 ri->ri_lsegs_start_seq = seg_seq;
 888                         }
 889                         if (flags & NILFS_SS_LOGEND)
 890                                 ri->ri_lsegs_end = pseg_start;
 891                         goto try_next_pseg;
 892                 }
 893
 894                 /* A valid super root was found. */
 895                 ri->ri_cno = cno++;
 896                 ri->ri_super_root = pseg_end;
 897                 ri->ri_lsegs_start = ri->ri_lsegs_end = 0;
 898
 899                 nilfs_dispose_segment_list(&segments);
 900                 sr_pseg_start = pseg_start;
 901                 nilfs->ns_pseg_offset = pseg_start + nblocks - seg_start;
 902                 nilfs->ns_seg_seq = seg_seq;
 903                 nilfs->ns_segnum = segnum;
 904                 nilfs->ns_cno = cno;  /* nilfs->ns_cno = ri->ri_cno + 1 */
 905                 nilfs->ns_ctime = le64_to_cpu(sum->ss_create);
 906                 nilfs->ns_nextnum = nextnum;
 907
 908                 if (scan_newer)
 909                         ri->ri_need_recovery = NILFS_RECOVERY_SR_UPDATED;
 910                 else {
 911                         if (nilfs->ns_mount_state & NILFS_VALID_FS)
 912                                 goto super_root_found;
 913                         scan_newer = 1;
 914                 }
 915
 916  try_next_pseg:
 917                 /* Standing on a course, or met an inconsistent state */
 918                 pseg_start += nblocks;
 919                 if (pseg_start < seg_end)
 920                         continue;
 921                 goto feed_segment;
 922
 923  strayed:
 924                 /* Off the trail */
 925                 if (!scan_newer)
 926                         /*
 927                          * This can happen if a checkpoint was written without
 928                          * barriers, or as a result of an I/O failure.
 929                          */
 930                         goto failed;
 931
 932  feed_segment:
 933                 /* Looking to the next full segment */
 934                 if (empty_seg++)
 935                         goto super_root_found; /* found a valid super root */
 936
 937                 ret = nilfs_segment_list_add(&segments, segnum);
 938                 if (unlikely(ret))
 939                         goto failed;
 940
 941                 seg_seq++;
 942                 segnum = nextnum;
 943                 nilfs_get_segment_range(nilfs, segnum, &seg_start, &seg_end);
 944                 pseg_start = seg_start;
 945         }
 946
 947  super_root_found:
 948         /* Updating pointers relating to the latest checkpoint */
 949         brelse(bh_sum);
 950         list_splice_tail(&segments, &ri->ri_used_segments);
 951         nilfs->ns_last_pseg = sr_pseg_start;
 952         nilfs->ns_last_seq = nilfs->ns_seg_seq;
 953         nilfs->ns_last_cno = ri->ri_cno;
 954         return 0;
 955
 956  failed:
 957         brelse(bh_sum);
 958         nilfs_dispose_segment_list(&segments);
 959         return (ret < 0) ? ret : nilfs_warn_segment_error(ret);
 960 }