From: Al Viro > Sent: 27 July 2020 14:48 > > On Mon, Jul 27, 2020 at 09:51:45AM +0000, David Laight wrote: > > > I'm sure there is code that processes options in chunks. > > This probably means it is possible to put a chunk boundary > > at the end of userspace and continue processing the very start > > of kernel memory. > > > > At best this faults on the kernel copy code and crashes the system. > > Really? Care to provide some details, or is it another of your "I can't > be possibly arsed to check what I'm saying, but it stands for reason > that..." specials? I did more 'homework' than sometimes :-) Slightly difficult without a searchable net-next tree. However, as has been pointed out is a different thread this code is used to update IPv6 flow labels: > > - if (copy_from_user(fl->opt+1, optval+CMSG_ALIGN(sizeof(*freq)), olen)) > > + sockptr_advance(optval, CMSG_ALIGN(sizeof(*freq))); > > + if (copy_from_sockptr(fl->opt + 1, optval, olen)) > > goto done; and doesn't work because the advances are no longer cumulative. Now access_ok() has to take the base address and length to stop 'running into' kernel space, but the code above can advance from a valid user pointer (which won't fault) to a kernel address. If there were always an unmapped 'guard' page in the user address space the access_ok() check prior to copy_to/from_user() wouldn't need the length. So I surmise that no such guard page exists and so the above can advance from user addresses into kernel ones. David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
