Uploaded image for project: 'Lustre'
  1. Lustre
  2. LU-21283

debugfs: logdump -O loops forever on external journal with s_start 0

XMLWordPrintable

    • Icon: Bug Bug
    • Resolution: Unresolved
    • Icon: Minor Minor
    • None
    • None
    • 3
    • 9223372036854775807

      debugfs logdump -O cannot dump an external journal device (or an image of one) whose journal superblock has s_start == 0, which is the normal state after e2fsck has recovered the journal. Seen while trying to get the list of blocks written by a journal replay from an image of an external journal. The code is the same in e2fsprogs master-lustre (1.47.3-wc4) and upstream master.

      Reproducer

      mke2fs -F -b 4096 -O journal_dev -U <uuid> jnl.img 1024
      mke2fs -F -b 4096 -O ^has_journal fs.img 8192
      debugfs -w -R "feature has_journal" fs.img      # + ssv journal_dev / journal_uuid
      debugfs -w -f - fs.img <<< $'jo -f jnl.img\njw -b 8000-8003 /dev/zero\njc'
      e2fsck -fy -j jnl.img fs.img                    # replays, sets s_start = 0
      debugfs -R "logdump -aO -f jnl.img" fs.img      # never terminates
      debugfs -R "logdump -S -f jnl.img"              # segfault
      

      Output repeats forever:

      Found sequence 0 (not 3) at block 1: end of journal.
      Found expected sequence 0, type 4 (V2 superblock) at block 1
      Unexpected block type 4 at block 1.
      

      Cause (debugfs/logdump.c, dump_journal())

      1. With s_start == 0 and -O, the dump starts at a hard-coded block 1. That is the first log block of an internal journal, but on an external journal device block 0 holds the ext2 superblock, block 1 the journal superblock, and the log starts at s_first (2 with 4 KiB blocks, 3 with 1 KiB blocks).
      2. In the default: case ("Unexpected block type") the break only leaves the switch, so the while loop re-reads the same block forever.
      3. logdump S -f <journal> without an open filesystem passes current_fs>blocksize to e2p_list_journal_super() with current_fs == NULL.

      Fix
      Start at s_first instead of 1, end the dump on an unexpected block type, and pass 0 as the expected block size when no filesystem is open. New test j_ext_logdump_old: it hangs and segfaults without the fix and passes with it. The existing j_ and f_jnl_ logdump tests still pass.

      Not fixed here: with -O the fast-commit dump reads one block past the end of a journal without a fast-commit area and prints "short read".

      Workaround for unpatched debugfs: copy the image and set s_start (journal superblock offset 0x1c) to s_first (offset 0x14), then run logdump -aO -n <large> -f <copy>.

            zam Alexander Zarochentsev
            zam Alexander Zarochentsev
            Votes:
            0 Vote for this issue
            Watchers:
            3 Start watching this issue

              Created:
              Updated: