Filesort uses Field::sort_length() to calculate space needed to sort the field.

In case of BLOB and TEXT fields and all their
TINY, MEDIUM, LONG variations the length is calculated as follows:

uint32 Field_blob::sort_length() const
{
return (uint32) (current_thd->variables.max_sort_length +
(field_charset == &my_charset_bin ? 0 : packlength));
}

The default value of max_sort_length is 1024.

This is bad for TINYBLOB and TINYTEXT.

• It should be enough to use 256 bytes to sort TINYBLOB,
• It should be enough to use 256*strxfrm_multiply bytes to sort TINYTEXT.
  (where strxfrm_multiply is 1 for many collations, which makes 256 bytes again)

So TINYBLOB (and TINYTEXT in most cases) use 4 times more space
for sorting than it's actually needed. That should affect performance
very negatively.