Re: [Xen-devel] [PATCH 3/4] xen/PVH: Set up GS segment for stack canary

[Jan Beulich]

>>> On 02.05.18 at 19:29, <boris.ostrovsky@xxxxxxxxxx> wrote:
> On 05/02/2018 11:41 AM, Jan Beulich wrote:
>>>>> On 02.05.18 at 17:22, <boris.ostrovsky@xxxxxxxxxx> wrote:
>>> On 05/02/2018 11:01 AM, Jan Beulich wrote:
>>>>>>> On 02.05.18 at 17:00, <boris.ostrovsky@xxxxxxxxxx> wrote:
>>>>> On 05/02/2018 04:16 AM, Jan Beulich wrote:
>>>>>>>>> On 30.04.18 at 18:23, <boris.ostrovsky@xxxxxxxxxx> wrote:
>>>>>>> --- a/arch/x86/xen/xen-pvh.S
>>>>>>> +++ b/arch/x86/xen/xen-pvh.S
>>>>>>> @@ -54,6 +54,9 @@
>>>>>>>   * charge of setting up it's own stack, GDT and IDT.
>>>>>>>   */
>>>>>>>  
>>>>>>> +#define PVH_GDT_ENTRY_CANARY    4
>>>>>>> +#define PVH_CANARY_SEL          (PVH_GDT_ENTRY_CANARY * 8)
>>>>>> I can only advise against doing it this way: There's no safeguard against
>>>>>> someone changing asm/segment.h without changing this value (in fact
>>>>>> this applies to all of the GDT selectors populated in this file). At the 
>>>>> very
>>>>>> least tie this to GDT_ENTRY_BOOT_TSS / __BOOT_TSS?
>>>>>>
>>>>>>> @@ -64,6 +67,9 @@ ENTRY(pvh_start_xen)
>>>>>>>  	mov %eax,%es
>>>>>>>  	mov %eax,%ss
>>>>>>>  
>>>>>>> +	mov $(PVH_CANARY_SEL),%eax
>>>>>>> +	mov %eax,%gs
>>>>>>> +
>>>>>>>  	/* Stash hvm_start_info. */
>>>>>>>  	mov $_pa(pvh_start_info), %edi
>>>>>>>  	mov %ebx, %esi
>>>>>>> @@ -150,6 +156,7 @@ gdt_start:
>>>>>>>  	.quad 0x00cf9a000000ffff            /* __BOOT_CS */
>>>>>>>  #endif
>>>>>>>  	.quad 0x00cf92000000ffff            /* __BOOT_DS */
>>>>>>> +	.quad 0x0040900000000018            /* PVH_CANARY_SEL */
>>>>>> Without any further code before loading the selector, this points at
>>>>>> physical address 0. Don't you need to add in the base address of
>>>>>> the per-CPU stack_canary?
>>>>> This GDT is gone soon after we jump into generic x86 startup code.That
>>>>> code will load its own GDT (and then set up per-cpu segments and all that).
>>>> All understood, but why would you set up the per-CPU segment here if
>>>> what you load into the segment register is not usable for the intended
>>>> purpose (until that other code sets up things and reloads the segment
>>>> registers)?
>>> The intended purpose here is to allow stack protector access not to
>>> fail. At this point it doesn't really matter that GS is later used for
>>> per-cpu segment, this code (and this GDT) will not be used when other
>>> CPUs come up.
>> But the place the canary would live this way is completely wrong. 
> 
> 
> Would creating a canary variable and using it as a base address be better?

Of course, because then at least you properly control where an eventual
access would go, instead of touching some unrelated memory location.

Jan