Re: [git pull] core, x86: make LIST_POISON less deadly

[Avi Kivity <avi@qumranet.com>]

Ingo Molnar wrote:
> * Linus Torvalds <torvalds@linux-foundation.org> wrote:
>
>   
>> On Mon, 14 Jul 2008, Ingo Molnar wrote:
>>     
>>>  
>>> +config ILLEGAL_POINTER_VALUE
>>> +       hex
>>> +       default 0 if X86_32
>>> +       default 0xffffc10000000000 if X86_64
>>>       
>> This looks like a singularly bad pointer value on x86-64.
>>
>> Why not pick something that is *guaranteed* to fault? The above looks 
>> like any future setup that supports 41 bits of addressing and has 
>> extended the page tables (yes, it will happen eventually) will find 
>> that to be a perfectly valid address?
>>
>> It's also visually confusing, since it's visually very close to a real 
>> kernel pointer too.
>>
>> Grr.
>>
>> Why not use something sane like 0xdead000000000000, which has the high 
>> bit set but very fundamentally isn't a valid pointer, and never will 
>> be? And which is a *lot* more visually obvious too!
>>     
>
> initially i suggested that too - but such addresses raise a #GP instead 
> of a page fault so their decoding is a bit harder.
>
> We dont do any instruction decoding in #GP handlers to figure out what 
> happened, while in the pagefault case we know which address faulted, 
> etc.
>
> Perhaps we could try to make #GP handlers a bit more informative - 
> although decoding instructions will make things a bit more fragile 
> inevitably.
>
> Perhaps make it 0xffffcdead0000000 ?
>   

We could have the oops handler detect this address range, and point out 
the problem in plain English.

-- 
error compiling committee.c: too many arguments to function