Re: [PATCHv3] x86/mm: set x32 syscall bit in SET_PERSONALITY()

[Dmitry Safonov]

2017-03-22 0:16 GMT+03:00 Adam Borowski <kilobyte@xxxxxxxxxx>:
> 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.

JFI: x32 mmap() syscall in 64 ELF should work even better - it has returned
addresses over 4Gb in ia32 mmap()s, so I expect it did the same in x32 top-down
allocation. (I guess you've mentioned the fixes-for patch).
So the thing to check not also that mmap() returned address, but at least
verify-dereference it with `mov' e.g. (or better - to parse /proc/self/maps)

> 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.

Big thanks for the testing work, Adam!

-- 
             Dmitry