Re: [Xen-devel] [PATCH 4/5] x86/pv: Drop support for paging out the LDT

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 18/01/18 10:57, Andrew Cooper wrote:
> On 18/01/18 10:38, George Dunlap wrote:
>> On Thu, Jan 18, 2018 at 7:59 AM, Jan Beulich <JBeulich@xxxxxxxx> wrote:
>>>>>> On 12.01.18 at 19:37, <andrew.cooper3@xxxxxxxxxx> wrote:
>>>> Windows is the only OS which pages out kernel datastructures, so chances 
>>>> are
>>>> good that this is a vestigial remnant of the PV Windows XP experiment.
>>> This is based on what? How do you know there are no other OSes
>>> doing so, including perhaps ones which none of us has ever heard
>>> of?
> The chances of there being a production OS ported to PV that we've never
> head of 0.
>
>>>  The diffstat of the change here is certainly nice, but as always
>>> with something being removed from the ABI I'm rather hesitant.
>>>
>>>> --- a/xen/arch/x86/domain.c
>>>> +++ b/xen/arch/x86/domain.c
>>>> @@ -1942,11 +1942,8 @@ int domain_relinquish_resources(struct domain *d)
>>>>          {
>>>>              for_each_vcpu ( d, v )
>>>>              {
>>>> -                /*
>>>> -                 * Relinquish GDT mappings. No need for explicit 
>>>> unmapping of
>>>> -                 * the LDT as it automatically gets squashed with the 
>>>> guest
>>>> -                 * mappings.
>>>> -                 */
>>>> +                /* Relinquish GDT/LDT mappings. */
>>>> +                pv_destroy_ldt(v);
>>>>                  pv_destroy_gdt(v);
>>> The domain is dead at this point, so the order doesn't matter much,
>>> but strictly speaking you should destroy the GDT before destroying
>>> the LDT (just like LDT _loads_ always need to come _after_ GDT
>>> adjustments).
> By that logic, I've got it the correct way round (which is they way I
> intended).  Allocation and freeing of the LDT is nested within the scope
> of the GDT.
>
>>> Everything else here looks fine, but the initial comment may need
>>> further discussion. For example we may want to consider a
>>> two-stage phasing out of the feature, with a couple of years in
>>> between: Make the functionality dependent upon a default-off
>>> command line option for the time being, and issue a bright warning
>>> when someone actually enables it (telling them to tell us).
>> One of the problems we have is that people seem to wait for 2-3 years
>> after a release has been made to start updating to it.  So we'd have
>> to leave such a warning for probably 5 years minimum.
> In almost any other case, I'd agree, and would be the first to suggest this.
>
> However, This patch is strictly necessary for a more complete XPTI,
> which is how I stumbled onto the issue in my KAISER series to begin
> with.  It is the only codepath where we ever poke at a remote critical
> data structure, which is why

Sorry - sent too early.

... which is why isolating the LDT in a per-cpu doesn't work in
combination with this algorithm.

~Andrew

>
> Also as noted in the commit message, it is broken even in the case you
> wanted to sensibly page the LDT, which further backs up the exception
> that it was only for Windows XP, and has never ever encountered a
> production system.
>
> ~Andrew


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel