[LU-15590] e2fsprogs direct IO raw_read_blk() issues

Related — Tue, 19 Jul 2022 23:25:12 +0000

In e2fsprogs (master=v1.46.2-9-g32fda1e5 and recent wc tags), there is a one line bug in the direct IO case of raw_read_blk().

bounce_read:
        if (channel->align == 0)
                channel->align = 1;
        if ((channel->block_size > channel->align) &&
            (channel->block_size % channel->align) == 0)
                align_size = channel->block_size;
        else
                align_size = channel->align;
        aligned_blk = location / align_size;
        offset = location % align_size;

        mutex_lock(data, BOUNCE_MTX);
        if (ext2fs_llseek(data->dev, aligned_blk * align_size, SEEK_SET) < 0) {
                retval = errno ? errno : EXT2_ET_LLSEEK_FAILED;
                goto error_unlock;
        }
        while (size > 0) {
                actual = read(data->dev, data->bounce, align_size);
                if (actual != align_size) {
                        mutex_unlock(data, BOUNCE_MTX);
                        actual = really_read;
                        buf -= really_read;
                        size += really_read;
                        goto short_read;
                }
                actual = size; /* HERE HIER AQUI ICI QUI TUK */
                if (actual > align_size)
                        actual = align_size;
                actual -= offset;
                memcpy(buf, data->bounce + offset, actual);

We should not be setting actual = size.

When this is called from ext2fs_open() to read the super block on devices with a 4KB logical block size and EXT2_FLAG_DIRECT_IO it mis-initializes the super block and ext2fs_open() fails with EXT2_ET_BAD_MAGIC.

This can be reproduced in a few ways.

Get a recent version of losetup that supports the -b option to set the logical blocksize.
Hack ext2fs_get_device_sectsize() to return 4K.
Use a device with 4K blocks.

Note, that you need to use EXT2_FLAG_DIRECT_IO in any of these cases to reproduce the issue.

It looks like this was fixed by artem_blagodarenko's change

c0015961 libext2fs: fix unix_io's Direct I/O support

and then inadvertently reintroduced by

d557b965 libext2fs: fix potential races in unix_io

See https://marc.info/?l=linux-ext4&m=161461569206170&w=2

Whamcloud Community JIRA

[LU-15590] e2fsprogs direct IO raw_read_blk() issues