[RFD] x86: The future of MPX

[RFD] x86: The future of MPX

[Thomas Gleixner]

Dave,

I've got the following MPX related information:

 .... But they [Intel] certainly haven't maintained the whole MPX support
 in GCC at all since two or three years (e.g. bootstrap with MPX was broken
 since forever), so now it's gone from GCC. ....

and I know that GCC is discussing MPX removal since May last year at least.
To the best of my knowledge clang does not support it either.

So what's the state of this stuff? If we don't have a compiler supporting
it how on earth is that useful or even testable for anyone who does not
have access to some magic Intel internal compiler version?

If that's the sad state of affairs, then we just should move the whole
stuff into staging or simply queue it for removal in 4.18.

Thanks,

	tglx

Re: [RFD] x86: The future of MPX

[Ingo Molnar]

* Thomas Gleixner <tglx@linutronix.de> wrote:

> Dave,
> 
> I've got the following MPX related information:
> 
>  .... But they [Intel] certainly haven't maintained the whole MPX support
>  in GCC at all since two or three years (e.g. bootstrap with MPX was broken
>  since forever), so now it's gone from GCC. ....
> 
> and I know that GCC is discussing MPX removal since May last year at least.
> To the best of my knowledge clang does not support it either.
> 
> So what's the state of this stuff? If we don't have a compiler supporting
> it how on earth is that useful or even testable for anyone who does not
> have access to some magic Intel internal compiler version?
> 
> If that's the sad state of affairs, then we just should move the whole
> stuff into staging or simply queue it for removal in 4.18.

I just tried the MPX testcases with the latest kernel:

 deimos:~/tip/tools/testing/selftests/x86> ./mpx-mini-test_64 
 XSAVE is supported by HW & OS
 XSAVE processor supported state mask: 0x21f
 XSAVE OS supported state mask: 0x21f
  BNDREGS: size: 64 user: 1 supervisor: 0 aligned: 0
   BNDCSR: size: 64 user: 1 supervisor: 0 aligned: 0
 mpx-mini-test_64: mpx-mini-test.c:58: write_int_to: Assertion `fd >= 0' failed.
 Aborted (core dumped)

 deimos:~/tip/tools/testing/selftests/x86> ./mpx-mini-test_32 
 XSAVE is supported by HW & OS
 XSAVE processor supported state mask: 0x21f
 XSAVE OS supported state mask: 0x21f
  BNDREGS: size: 64 user: 1 supervisor: 0 aligned: 0
   BNDCSR: size: 64 user: 1 supervisor: 0 aligned: 0
 mpx-mini-test_32: mpx-mini-test.c:58: write_int_to: Assertion `fd >= 0' failed.
 Aborted (core dumped)

... and both of them crashed.

The CPU has MPX support and the kernel has CONFIG_X86_INTEL_MPX=y:

 deimos:~> dmesg | grep -i mpx
 [    0.000000] x86/fpu: Supporting XSAVE feature 0x008: 'MPX bounds registers'
 [    0.000000] x86/fpu: Supporting XSAVE feature 0x010: 'MPX CSR'

Thanks,

	Ingo

Re: [RFD] x86: The future of MPX

[Linus Torvalds]

On Sat, Apr 28, 2018 at 2:18 AM Ingo Molnar <mingo@kernel.org> wrote:

> I just tried the MPX testcases with the latest kernel:

>   deimos:~/tip/tools/testing/selftests/x86> ./mpx-mini-test_64

They need to be run as root. They try to access

     /sys/kernel/debug/tracing/set_ftrace_pid
     /sys/kernel/debug/tracing/trace

and that's root-only.

> ... and both of them crashed.

That's just because they use "assert()" to check that the file open worked.

"assert()" is evil and bad bad bad.

                      Linus

Re: [RFD] x86: The future of MPX

[Linus Torvalds]

On Sat, Apr 28, 2018 at 9:36 AM Linus Torvalds <
torvalds@linux-foundation.org> wrote:

> They need to be run as root.

Side note: don't get me wrong. If the MPX stuff isn't supported by gcc,
then there is little point in us supporting it in the kernel either.

So I'm just saying that the test-case is root-only and the crash doesn't
necessarily mean anything else than "badly written test".

                   Linus

Re: [RFD] x86: The future of MPX

[Ingo Molnar]

* Linus Torvalds <torvalds@linux-foundation.org> wrote:

> On Sat, Apr 28, 2018 at 2:18 AM Ingo Molnar <mingo@kernel.org> wrote:
> 
> > I just tried the MPX testcases with the latest kernel:
> 
> >   deimos:~/tip/tools/testing/selftests/x86> ./mpx-mini-test_64
> 
> They need to be run as root. They try to access
> 
>      /sys/kernel/debug/tracing/set_ftrace_pid
>      /sys/kernel/debug/tracing/trace
> 
> and that's root-only.

Indeed - but note that they are crashing as root as well here:

  root@deimos:/home/mingo/tip/tools/testing/selftests/x86# ./mpx-mini-test_64 
  [...]
  Aborted (core dumped)

... because I don't even have /sys/kernel/debug/tracing/set_ftrace_pid... :-/

That's despite having ftrace enabled.

After some digging I found out that 'set_ftrace_pid' is dependent on 
CONFIG_FUNCTION_TRACING=y (not just CONFIG_TRACING=y), which I didn't
have enabled on this kernel.

After enabling it I got the 64-bit testcase to work:

  #
  ...
  ./mpx-mini-test_64 completed successfully

But the 32-bit testcase is erroring out:

  root@deimos:/home/mingo/tip/tools/testing/selftests/x86# ./mpx-mini-test_32 
  XSAVE is supported by HW & OS
  XSAVE processor supported state mask: 0x21f
  XSAVE OS supported state mask: 0x21f
   BNDREGS: size: 64 user: 1 supervisor: 0 aligned: 0
    BNDCSR: size: 64 user: 1 supervisor: 0 aligned: 0
  executing unmaptest
  mpx dig (  1) complete, SUCCESS (       0 /    0)
  #BR status == 2, missing bounds table,kernel should have handled!!

v4.17-rc2 based kernel, so it should have the latest MPX code.

Thanks,

	Ingo

Re: [RFD] x86: The future of MPX

[Eric W. Biederman]

Ingo Molnar <mingo@kernel.org> writes:

> * Linus Torvalds <torvalds@linux-foundation.org> wrote:
>
>> On Sat, Apr 28, 2018 at 2:18 AM Ingo Molnar <mingo@kernel.org> wrote:
>> 
>> > I just tried the MPX testcases with the latest kernel:
>> 
>> >   deimos:~/tip/tools/testing/selftests/x86> ./mpx-mini-test_64
>> 
>> They need to be run as root. They try to access
>> 
>>      /sys/kernel/debug/tracing/set_ftrace_pid
>>      /sys/kernel/debug/tracing/trace
>> 
>> and that's root-only.
>
> Indeed - but note that they are crashing as root as well here:
>
>   root@deimos:/home/mingo/tip/tools/testing/selftests/x86# ./mpx-mini-test_64 
>   [...]
>   Aborted (core dumped)
>
> ... because I don't even have /sys/kernel/debug/tracing/set_ftrace_pid... :-/
>
> That's despite having ftrace enabled.
>
> After some digging I found out that 'set_ftrace_pid' is dependent on 
> CONFIG_FUNCTION_TRACING=y (not just CONFIG_TRACING=y), which I didn't
> have enabled on this kernel.
>
> After enabling it I got the 64-bit testcase to work:
>
>   #
>   ...
>   ./mpx-mini-test_64 completed successfully
>
> But the 32-bit testcase is erroring out:
>
>   root@deimos:/home/mingo/tip/tools/testing/selftests/x86# ./mpx-mini-test_32 
>   XSAVE is supported by HW & OS
>   XSAVE processor supported state mask: 0x21f
>   XSAVE OS supported state mask: 0x21f
>    BNDREGS: size: 64 user: 1 supervisor: 0 aligned: 0
>     BNDCSR: size: 64 user: 1 supervisor: 0 aligned: 0
>   executing unmaptest
>   mpx dig (  1) complete, SUCCESS (       0 /    0)
>   #BR status == 2, missing bounds table,kernel should have handled!!
>
> v4.17-rc2 based kernel, so it should have the latest MPX code.

If you remove this which tree will this be going into?

I have some refactoring and cleanups of the x86 siginfo code I am
hoping to get in later this development cycle and part of that was
refactoring the mpx code so it works more like everything else.

If the codes away I can just that part of the patchset.

Eric

Re: [RFD] x86: The future of MPX

[Dave Hansen]

On 04/28/2018 09:44 AM, Linus Torvalds wrote:
> On Sat, Apr 28, 2018 at 9:36 AM Linus Torvalds <
> torvalds@linux-foundation.org> wrote:
> 
>> They need to be run as root.
> Side note: don't get me wrong. If the MPX stuff isn't supported by gcc,
> then there is little point in us supporting it in the kernel either.

The loss of the GCC support is definitely a bummer.

There are no objections from me against removing it from future kernels.
 I think it will be pretty straightforward, just leaving the prctl()
numbers as placeholders.

Re: [RFD] x86: The future of MPX

[Borislav Petkov]

On Sun, Apr 29, 2018 at 09:44:00PM -0700, Dave Hansen wrote:
> The loss of the GCC support is definitely a bummer.

https://gcc.gnu.org/ml/gcc-patches/2018-05/msg00027.html

Will go into gcc9.

-- 
Regards/Gruss,
    Boris.

Good mailing practices for 400: avoid top-posting and trim the reply.

Re: [RFD] x86: The future of MPX

[Eric W. Biederman]

Borislav Petkov <bp@alien8.de> writes:

> On Sun, Apr 29, 2018 at 09:44:00PM -0700, Dave Hansen wrote:
>> The loss of the GCC support is definitely a bummer.
>
> https://gcc.gnu.org/ml/gcc-patches/2018-05/msg00027.html
>
> Will go into gcc9.

So we should see this in a released version of gcc in about a year?

Eric