On Tue, Mar 21, 2017 at 08:47:11PM +0300, Dmitry Safonov wrote:
> After my changes to mmap(), its code now relies on the bitness of
> performing syscall. According to that, it chooses the base of allocation:
> mmap_base for 64-bit mmap() and mmap_compat_base for 32-bit syscall.
> It was done by:
>   commit 1b028f784e8c ("x86/mm: Introduce mmap_compat_base() for
> 32-bit mmap()").
> 
> The code afterwards relies on in_compat_syscall() returning true for
> 32-bit syscalls. It's usually so while we're in context of application
> that does 32-bit syscalls. But during exec() it is not valid for x32 ELF.
> The reason is that the application hasn't yet done any syscall, so x32
> bit has not being set.
> That results in -ENOMEM for x32 ELF files as there fired BAD_ADDR()
> in elf_map(), that is called from do_execve()->load_elf_binary().
> For i386 ELFs it works as SET_PERSONALITY() sets TS_COMPAT flag.
> 
> Set x32 bit before first return to userspace, during setting personality
> at exec(). This way we can rely on in_compat_syscall() during exec().
> Do also the reverse: drop x32 syscall bit at SET_PERSONALITY for 64-bits.
> 
> Fixes: commit 1b028f784e8c ("x86/mm: Introduce mmap_compat_base() for
> 32-bit mmap()")

Tested:
with bash:x32, mksh:amd64, posh:i386, zsh:armhf (binfmt:qemu), fork+exec
works for every parent-child combination.

Contrary to my naive initial reading of your fix, mixing syscalls from a
process of the wrong ABI also works as it did before.  While using a glibc
wrapper will call the right version, x32 processes calling amd64 syscalls is
surprisingly common -- this brings seccomp joy.

I've attached a freestanding test case for write() and mmap(); it's
freestanding asm as most of you don't have an x32 toolchain at hand, sorry
for unfriendly error messages.

So with these two patches:
x86/tls: Forcibly set the accessed bit in TLS segments
x86/mm: set x32 syscall bit in SET_PERSONALITY()
everything appears to be fine.

-- 
ac?aGBP'a  3/4 a >>ac?aGBP|a ? Meow!
aGBP 3/4 a ?ac a ?a ?aGBP?a!?
ac?a!?a ?a .a ?a ?a ? Collisions shmolisions, let's see them find a collision or second
a ?a 3aGBP?a ?a ?a ?a ? preimage for double rot13!