Re: [RFC v2 7/9] Trampoline emulation

[Salvatore Mesoraca]

2017-06-28 1:13 GMT+02:00 Kees Cook <keescook@xxxxxxxxxxxx>:
> On Thu, Jun 15, 2017 at 9:42 AM, Salvatore Mesoraca
> <s.mesoraca16@xxxxxxxxx> wrote:
>> Some programs need to generate part of their code at runtime. Luckily
>> enough, in some cases they only generate well-known code sequences (the
>> "trampolines") that can be easily recognized and emulated by the kernel.
>> This way WX Protection can still be active, so a potential attacker won't
>> be able to generate arbitrary sequences of code, but just those that are
>> explicitly allowed. This is not ideal, but it's still better than having WX
>> Protection completely disabled.
>> In particular S.A.R.A. is able to recognize trampolines used by GCC for
>> nested C functions and libffi's trampolines.
>> This feature is implemented only on x86_32 and x86_64.
>> The assembly sequences used here were originally obtained from PaX source
>> code.
>
> See below about the language grsecurity has asked people to use in commit logs.

OK, I'll change the commit message in v3.

>>
>> Signed-off-by: Salvatore Mesoraca <s.mesoraca16@xxxxxxxxx>
>> ---
>>  security/sara/Kconfig               |  17 ++++
>>  security/sara/include/trampolines.h | 171 ++++++++++++++++++++++++++++++++++++
>>  security/sara/wxprot.c              | 140 +++++++++++++++++++++++++++++
>>  3 files changed, 328 insertions(+)
>>  create mode 100644 security/sara/include/trampolines.h
>>
>> diff --git a/security/sara/Kconfig b/security/sara/Kconfig
>> index 6c74069..f406805 100644
>> --- a/security/sara/Kconfig
>> +++ b/security/sara/Kconfig
>> @@ -96,6 +96,23 @@ choice
>>                   Documentation/security/SARA.rst.
>>  endchoice
>>
>> +config SECURITY_SARA_WXPROT_EMUTRAMP
>> +       bool "Enable emulation for some types of trampolines"
>> +       depends on SECURITY_SARA_WXPROT
>> +       depends on X86
>> +       default y
>> +       help
>> +         Some programs and libraries need to execute special small code
>> +         snippets from non-executable memory pages.
>> +         Most notable examples are the GCC and libffi trampolines.
>> +         This features make it possible to execute those trampolines even
>> +         if they reside in non-executable memory pages.
>> +         This features need to be enabled on a per-executable basis
>> +         via user-space utilities.
>> +         See Documentation/security/SARA.rst. for further information.
>> +
>> +         If unsure, answer y.
>> +
>>  config SECURITY_SARA_WXPROT_DISABLED
>>         bool "WX protection will be disabled at boot."
>>         depends on SECURITY_SARA_WXPROT
>> diff --git a/security/sara/include/trampolines.h b/security/sara/include/trampolines.h
>> new file mode 100644
>> index 0000000..eab0a85
>> --- /dev/null
>> +++ b/security/sara/include/trampolines.h
>> @@ -0,0 +1,171 @@
>> +/*
>> + * S.A.R.A. Linux Security Module
>> + *
>> + * Copyright (C) 2017 Salvatore Mesoraca <s.mesoraca16@xxxxxxxxx>
>> + *
>> + * This program is free software; you can redistribute it and/or modify
>> + * it under the terms of the GNU General Public License version 2, as
>> + * published by the Free Software Foundation.
>> + *
>> + * Assembly sequences used here were copied from
>> + * PaX patch by PaX Team <pageexec@xxxxxxxxxxx>
>
> Given this copying, please include the grsecurity/PaX copyright notice
> too. Please see the recommendations here:
> http://kernsec.org/wiki/index.php/Kernel_Self_Protection_Project/Get_Involved

I understand your concern, but I don't think that the formula
"Copyright (C) 2001-2017 PaX Team, Bradley Spengler, Open Source Security Inc"
is appropriate in this context.
I don't think that any of the code written here is covered by Grsecurity's
copyright. That line I wrote about "Assembly sequences" was just a courtesy
note because I didn't gdb-ed my way through all of them, but if someone owns
copyright on those "Assembly sequences" it's probably GCC and libffi authors.
I have no problem in giving the most appropriate credits to the relevant
people. And, for sure, I'm not going to get rich because of the attribution
of this code.
I wrote this code without copying any actual line and I prefer to not give
away my copyright without a good reason (let's call it a matter of principle).
I don't want to start the umpteenth flame war about grsecurity and copyright.
I'm acting in good faith and I hope this discussion can continue with
constructive input from the interested parties.

>> + *
>> + */
>> +
>> +#ifndef __SARA_TRAMPOLINES_H
>> +#define __SARA_TRAMPOLINES_H
>> +#ifdef CONFIG_SECURITY_SARA_WXPROT_EMUTRAMP
>> +
>> +
>> +/* x86_32 */
>> +
>> +
>> +struct libffi_trampoline_x86_32 {
>> +       unsigned char mov;
>> +       unsigned int addr1;
>> +       unsigned char jmp;
>> +       unsigned int addr2;
>> +} __packed;
>> +
>> +struct gcc_trampoline_x86_32_type1 {
>> +       unsigned char mov1;
>> +       unsigned int addr1;
>> +       unsigned char mov2;
>> +       unsigned int addr2;
>> +       unsigned short jmp;
>> +} __packed;
>> +
>> +struct gcc_trampoline_x86_32_type2 {
>> +       unsigned char mov;
>> +       unsigned int addr1;
>> +       unsigned char jmp;
>> +       unsigned int addr2;
>> +} __packed;
>> +
>> +union trampolines_x86_32 {
>> +       struct libffi_trampoline_x86_32 lf;
>> +       struct gcc_trampoline_x86_32_type1 g1;
>> +       struct gcc_trampoline_x86_32_type2 g2;
>> +};
>> +
>> +#define is_valid_libffi_trampoline_x86_32(UNION)       \
>> +       (UNION.lf.mov == 0xB8 &&                        \
>> +       UNION.lf.jmp == 0xE9)
>> +
>> +#define emulate_libffi_trampoline_x86_32(UNION, REGS) do {     \
>> +       (REGS)->ax = UNION.lf.addr1;                            \
>> +       (REGS)->ip = (unsigned int) ((REGS)->ip +               \
>> +                                    UNION.lf.addr2 +           \
>> +                                    sizeof(UNION.lf));         \
>> +} while (0)
>> +
>> +#define is_valid_gcc_trampoline_x86_32_type1(UNION, REGS)      \
>> +       (UNION.g1.mov1 == 0xB9 &&                               \
>> +       UNION.g1.mov2 == 0xB8 &&                                \
>> +       UNION.g1.jmp == 0xE0FF &&                               \
>> +       REGS->ip > REGS->sp)
>> +
>> +#define emulate_gcc_trampoline_x86_32_type1(UNION, REGS) do {  \
>> +       (REGS)->cx = UNION.g1.addr1;                            \
>> +       (REGS)->ax = UNION.g1.addr2;                            \
>> +       (REGS)->ip = UNION.g1.addr2;                            \
>> +} while (0)
>> +
>> +#define is_valid_gcc_trampoline_x86_32_type2(UNION, REGS)      \
>> +       (UNION.g2.mov == 0xB9 &&                                \
>> +       UNION.g2.jmp == 0xE9 &&                                 \
>> +       REGS->ip > REGS->sp)
>> +
>> +#define emulate_gcc_trampoline_x86_32_type2(UNION, REGS) do {  \
>> +       (REGS)->cx = UNION.g2.addr1;                            \
>> +       (REGS)->ip = (unsigned int) ((REGS)->ip +               \
>> +                                    UNION.g2.addr2 +           \
>> +                                    sizeof(UNION.g2));         \
>> +} while (0)
>
> These all seem like they need to live in arch/x86/... somewhere rather
> than in the LSM, but maybe this isn't needed on other architectures?
> This seems to be very arch and compiler specific...

It is very arch and compiler specific. At some extent it can even be
considered program specific, given that some of the trampolines are
used only by libffi (AFAIK).
I put it in S.A.R.A.'s directory because I didn't want to pollute
arch/x86 with code needed only by a single LSM.
In theory this feature could be useful on other archs too, but I'm not
sure if there is actual demand for it.
Do you have any suggestion about the best place in "arch/x86/..."
where it should be moved to?
Thank you very much for your time.

Salvatore