Kees Cook a écrit :
> Hi,
> 
> On Thu, Jan 27, 2011 at 06:24:14PM -0800, H. Peter Anvin wrote:
>> On 01/27/2011 03:00 PM, Kees Cook wrote:
>>> Yikes, good catch.
>>>
>>> arch/x86/kernel/trampoline_64.S uses:
>>>         movw    $(trampoline_stack_end - r_base), %sp
>>>
>>> arch/x86/boot/compressed/head_64.S uses:
>>>         movl    $boot_stack_end, %eax
>>>         addl    %ebp, %eax
>>>         movl    %eax, %esp
>>>
>>> what would be safe for arch/x86/kernel/head_32.S ? It uses "stack_start",
>>> but later after paging set-up. Is the following sane to solve this?
>>>
>> To run it before paging is set up, you can't use stack, start; you
>> have to use a pointer based on physical address.  You have two
>> problems with using stack_start: you're using a linear address to
>> access stack_start, and stack_start itself contains a linear
>> address.
>>
>> It's not entirely clear to me why we don't initialize %ss to
>> __BOOT_DS with the other segment registers, but it would make most
>> sense to me:
>>
>> diff --git a/arch/x86/kernel/head_32.S b/arch/x86/kernel/head_32.S
>> index fc293dc..c10f9ba 100644
>> --- a/arch/x86/kernel/head_32.S
>> +++ b/arch/x86/kernel/head_32.S
>> @@ -99,7 +99,12 @@ ENTRY(startup_32)
>>         movl %eax,%es
>>         movl %eax,%fs
>>         movl %eax,%gs
>> +       movl %eax,%ss
>>  2:
>> +/*
>> + * Set up an initial stack
>> + */
>> +       movl $pa(init_thread_union+THREAD_SIZE), %esp
>>
>>  /*
>>   * Clear BSS first so that there are no surprises...
> 
> This doesn't appear to work for me. While I can boot fine, doing CPU
> hotplugging hangs the system. :(
> 
This is weird because the patch only touch first cpu (startup_32 entry) and cpu hotplug go to 
startup_32_smp.

Here a untested patch that move the stack setup in the common path.


Matthieu