lkml.org 
[lkml]   [2008]   [Jul]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC 00/15] x86_64: Optimize percpu accesses
Mike Travis wrote:
> Eric W. Biederman wrote:
>
>> I just took a quick look at how stack_protector works on x86_64. Unless there is
>> some deep kernel magic that changes the segment register to %gs from the ABI specified
>> %fs CC_STACKPROTECTOR is totally broken on x86_64. We access our pda through %gs.
>>
>> Further -fstack-protector-all only seems to detect against buffer overflows and
>> thus corruption of the stack. Not stack overflows. So it doesn't appear especially
>> useful.
>>
>> So we don't we kill the broken CONFIG_CC_STACKPROTECTOR. Stop trying to figure out
>> how to use a zero based percpu area.
>>
>> That should allow us to make the current pda a per cpu variable, and use %gs with
>> a large offset to access the per cpu area. And since it is only the per cpu accesses
>> and the pda accesses that will change we should not need to fight toolchain issues
>> and other weirdness. The linked binary can remain the same.
>>
>> Eric
>>
>
> Hi Eric,
>
> There is one pda op that I was not able to remove. Most likely it can be recoded
> but it was a bit over my expertise. Most likely the "pda_offset(field)" can be
> replaced with "per_cpu_var(field)" [per_cpu__##field], but for "_proxy_pda.field"
> I wasn't sure about.
>
> include/asm-x86/pda.h:
>
> /*
> * This is not atomic against other CPUs -- CPU preemption needs to be off
> * NOTE: This relies on the fact that the cpu_pda is the *first* field in
> * the per cpu area. Move it and you'll need to change this.
> */
> #define test_and_clear_bit_pda(bit, field) \
> ({ \
> int old__; \
> asm volatile("btr %2,%%gs:%c3\n\tsbbl %0,%0" \
> : "=r" (old__), "+m" (_proxy_pda.field) \
> : "dIr" (bit), "i" (pda_offset(field)) : "memory");\
>
asm volatile("btr %2,%%gs:%1\n\tsbbl %0,%0" \
: "=r" (old__), "+m" (per_cpu_var(var)) \
: "dIr" (bit) : "memory");\


but it barely seems worthwhile if we really can't use test_and_clear_bit.

J


\
 
 \ /
  Last update: 2008-07-09 23:51    [W:0.224 / U:6.324 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site