Re: [RFC, PATCHv1 16/28] x86/asm: remove __VIRTUAL_MASK_SHIFT==47 assert

From: Andy Lutomirski
Date: Thu Dec 08 2016 - 13:40:28 EST

On Thu, Dec 8, 2016 at 8:21 AM, Kirill A. Shutemov
<kirill.shutemov@xxxxxxxxxxxxxxx> wrote:
> We don't need it anymore. 17be0aec74fb ("x86/asm/entry/64: Implement
> better check for canonical addresses") made canonical address check
> generic wrt. address width.

This code existed in part to remind us that this needs very careful
adjustment when the paging size becomes dynamic. If you want to
remove it, please add test cases to tools/testing/selftests/x86 that

a. Either mmap(2^47-4096, ..., MAP_FIXED, ...) fails or that, if it
succeeds and you put a syscall instruction at the very end, that
invoking the syscall instruction there works. The easiest way to do
this may be to have the selftest literally have a page of text that
has 4094 0xcc bytes and a syscall and to map that page or perhaps move
it into place with mremap. That will avoid annoying W^X userspace
stuff from messing up the test. You'll need to handle the signal when
you fall off the end of the world after the syscall.

b. Ditto for the new highest possible userspace page.

c. Ditto for one page earlier to make sure that your test actually works.

d. For each possible maximum address, call raise(SIGUSR1) and, in the
signal handler, change RIP to point to the first noncanonical address
and RCX to match RIP. Return and catch the resulting exception. This
may be easy to integrate into the sigreturn tests, and I can help with