Re: [PATCH v2 4/6] xen/x86/pvh: Set up descriptors for 32-bit PVH guests

[Boris Ostrovsky <boris.ostrovsky@oracle.com>]

On 07/17/2015 01:52 PM, Boris Ostrovsky wrote:
> On 07/17/2015 12:43 PM, Konrad Rzeszutek Wilk wrote:
>> On Fri, Jul 17, 2015 at 11:36:29AM -0400, Boris Ostrovsky wrote:
>>> On 07/17/2015 11:21 AM, Konrad Rzeszutek Wilk wrote:
>>>> On Thu, Jul 16, 2015 at 05:43:39PM -0400, Boris Ostrovsky wrote:
>>>>> Signed-off-by: Boris Ostrovsky <boris.ostrovsky@oracle.com>
>>>>> ---
>>>>> Changes in v2:
>>>>> * Set segment selectors using loadsegment() instead of assembly
>>>>>
>>>>>   arch/x86/xen/enlighten.c | 15 ++++++++++-----
>>>>>   1 file changed, 10 insertions(+), 5 deletions(-)
>>>>>
>>>>> diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
>>>>> index f8dc398..d665b1d 100644
>>>>> --- a/arch/x86/xen/enlighten.c
>>>>> +++ b/arch/x86/xen/enlighten.c
>>>>> @@ -1362,12 +1362,12 @@ static void __init 
>>>>> xen_boot_params_init_edd(void)
>>>>>   static void __ref xen_setup_gdt(int cpu)
>>>>>   {
>>>>>       if (xen_feature(XENFEAT_auto_translated_physmap)) {
>>>>> -#ifdef CONFIG_X86_64
>>>>> -        unsigned long dummy;
>>>>> +        unsigned long __attribute__((unused)) dummy;
>>>>> -        load_percpu_segment(cpu); /* We need to access per-cpu 
>>>>> area */
>>>> You removed that - where are we going to do that? As the
>>>> 'switch_to_new_gdt' uses the per-cpu GDT table.
>>> load_percpu_segment() is part of switch_to_new_gdt(), so I thought 
>>> there is
>>> no need to call it here.
>>>
>>> But you are right --- switch_to_new_gdt() starts with 
>>> get_cpu_gdt_table()
>>> which accesses per-CPU area. How did this manage to work then?
>> I was surprised as well - I was expecting your patch to have blow up.
>> Unless we are doing something fancy for CPU0 and for the other CPUs we
>> already have the per-cpu segment setup during bootup (copied from BSP)?
>
>
> No, %fs is zero when we enter xen_setup_gdt() (for 32-bit).
>
> In any case, I should put load_percpu_segment() back.


No, I shouldn't.

Until the new GDT is loaded we can't load selectors since current GDT 
doesn't have descriptors set up for them. And so any attempt to load 
uninitialized selectors results in a fault.

This worked for 64-bit guests because there we load zero into %gs and 
that is allowed (processor doesn't perform descriptor checks for the 
first 4 indexes). But for 32-bit guests we load %fs with 0xd8.

And the reason the code worked before was because we are using "master" 
per-cpu area and because GDT is the same for all CPUs at that point. Or 
so I think.


-boris