lkml.org 
[lkml]   [2015]   [Mar]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [BUG/PATCH] kernel RNG and its secrets
Date
Am Mittwoch, 18. März 2015, 13:19:07 schrieb Hannes Frederic Sowa:

Hi Hannes,

>On Wed, Mar 18, 2015, at 13:14, Stephan Mueller wrote:
>> Am Mittwoch, 18. März 2015, 13:02:12 schrieb Hannes Frederic Sowa:
>>
>> Hi Hannes,
>>
>> >On Wed, Mar 18, 2015, at 12:09, Stephan Mueller wrote:
>> >> Am Mittwoch, 18. März 2015, 11:56:43 schrieb Daniel Borkmann:
>> >> >On 03/18/2015 11:50 AM, Hannes Frederic Sowa wrote:
>> >> >> On Wed, Mar 18, 2015, at 10:53, mancha wrote:
>> >> >>> Hi.
>> >> >>>
>> >> >>> The kernel RNG introduced memzero_explicit in d4c5efdb9777 to
>> >> >>> protect
>> >> >>>
>> >> >>> memory cleansing against things like dead store optimization:
>> >> >>> void memzero_explicit(void *s, size_t count)
>> >> >>> {
>> >> >>>
>> >> >>> memset(s, 0, count);
>> >> >>> OPTIMIZER_HIDE_VAR(s);
>> >> >>>
>> >> >>> }
>> >> >>>
>> >> >>> OPTIMIZER_HIDE_VAR, introduced in fe8c8a126806 to protect
>> >> >>> crypto_memneq>>
>> >> >>>
>> >> >>> against timing analysis, is defined when using gcc as:
>> >> >>> #define OPTIMIZER_HIDE_VAR(var) __asm__ ("" : "=r" (var) :
>> >> >>> "0"
>> >> >>> (var))
>> >> >>>
>> >> >>> My tests with gcc 4.8.2 on x86 find it insufficient to prevent
>> >> >>> gcc
>> >> >>> from optimizing out memset (i.e. secrets remain in memory).
>> >> >>>
>> >> >>> Two things that do work:
>> >> >>> __asm__ __volatile__ ("" : "=r" (var) : "0" (var))
>> >> >>
>> >> >> You are correct, volatile signature should be added to
>> >> >> OPTIMIZER_HIDE_VAR. Because we use an output variable "=r", gcc
>> >> >> is
>> >> >> allowed to check if it is needed and may remove the asm
>> >> >> statement.
>> >> >> Another option would be to just use var as an input variable -
>> >> >> asm
>> >> >> blocks without output variables are always considered being
>> >> >> volatile
>> >> >> by gcc.
>> >> >>
>> >> >> Can you send a patch?
>> >> >>
>> >> >> I don't think it is security critical, as Daniel pointed out,
>> >> >> the
>> >> >> call
>> >> >> will happen because the function is an external call to the
>> >> >> crypto
>> >> >> functions, thus the compiler has to flush memory on return.
>> >> >
>> >> >Just had a look.
>> >> >
>> >> >$ gdb vmlinux
>> >> >(gdb) disassemble memzero_explicit
>> >> >
>> >> >Dump of assembler code for function memzero_explicit:
>> >> > 0xffffffff813a18b0 <+0>: push %rbp
>> >> > 0xffffffff813a18b1 <+1>: mov %rsi,%rdx
>> >> > 0xffffffff813a18b4 <+4>: xor %esi,%esi
>> >> > 0xffffffff813a18b6 <+6>: mov %rsp,%rbp
>> >> > 0xffffffff813a18b9 <+9>: callq 0xffffffff813a7120
>>
>> <memset>
>>
>> >> > 0xffffffff813a18be <+14>: pop %rbp
>> >> > 0xffffffff813a18bf <+15>: retq
>> >> >
>> >> >End of assembler dump.
>> >> >
>> >> >(gdb) disassemble extract_entropy
>> >> >[...]
>> >> >
>> >> > 0xffffffff814a5000 <+304>: sub %r15,%rbx
>> >> > 0xffffffff814a5003 <+307>: jne 0xffffffff814a4f80
>> >> >
>> >> ><extract_entropy+176> 0xffffffff814a5009 <+313>: mov %r12,%rdi
>> >> >
>> >> > 0xffffffff814a500c <+316>: mov $0xa,%esi
>> >> > 0xffffffff814a5011 <+321>: callq 0xffffffff813a18b0
>> >> >
>> >> ><memzero_explicit> 0xffffffff814a5016 <+326>: mov
>> >> >-0x48(%rbp),%rax
>> >> >[...]
>> >> >
>> >> >I would be fine with __volatile__.
>> >>
>> >> Are we sure that simply adding a __volatile__ works in any case? I
>> >> just did a test with a simple user space app:
>> >>
>> >> static inline void memset_secure(void *s, int c, size_t n)
>> >> {
>> >>
>> >> memset(s, c, n);
>> >> //__asm__ __volatile__("": : :"memory");
>> >> __asm__ __volatile__("" : "=r" (s) : "0" (s));
>> >>
>> >> }
>> >
>> >Good point, thanks!
>> >
>> >Of course an input or output of s does not force the memory pointed
>> >to
>> >by s being flushed.
>> >
>> >
>> >My proposal would be to add a
>> >
>> >#define OPTIMIZER_HIDE_MEM(ptr, len) __asm__ __volatile__ ("" : :
>> >"m"(
>> >({ struct { u8 b[len]; } *p = (void *)ptr ; *p; }) )
>> >
>> >and use this in the code function.
>> >
>> >This is documented in gcc manual 6.43.2.5.
>>
>> That one adds the zeroization instructuctions. But now there are much
>> more than with the barrier.
>>
>> 400469: 48 c7 04 24 00 00 00 movq $0x0,(%rsp)
>> 400470: 00
>> 400471: 48 c7 44 24 08 00 00 movq $0x0,0x8(%rsp)
>> 400478: 00 00
>> 40047a: c7 44 24 10 00 00 00 movl $0x0,0x10(%rsp)
>> 400481: 00
>> 400482: 48 c7 44 24 20 00 00 movq $0x0,0x20(%rsp)
>> 400489: 00 00
>> 40048b: 48 c7 44 24 28 00 00 movq $0x0,0x28(%rsp)
>> 400492: 00 00
>> 400494: c7 44 24 30 00 00 00 movl $0x0,0x30(%rsp)
>> 40049b: 00
>>
>> Any ideas?
>
>Hmm, correct definition of u8?

I use unsigned char
>
>Which version of gcc do you use? I can't see any difference if I
>compile your example at -O2.

gcc-Version 4.9.2 20150212 (Red Hat 4.9.2-6) (GCC)
>
>Bye,
>Hannes


Ciao
Stephan


\
 
 \ /
  Last update: 2015-03-18 13:41    [W:0.085 / U:0.188 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site