Re: [PATCH 42/43] x86/mm/kaiser: Allow KAISER to be enabled/disabled at runtime

[Andy Lutomirski]

> On Nov 25, 2017, at 1:05 PM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> 
> On Sat, 25 Nov 2017, Andy Lutomirski wrote:
>>>> On Nov 25, 2017, at 12:18 PM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>>>> On Fri, 24 Nov 2017, Ingo Molnar wrote:
>>>> 
>>>> From: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
>>>> 
>>>> The KAISER CR3 switches are expensive for many reasons.  Not all systems
>>>> benefit from the protection provided by KAISER.  Some of them can not
>>>> pay the high performance cost.
>>>> 
>>>> This patch adds a debugfs file.  To disable KAISER, you do:
>>>> 
>>>>   echo 0 > /sys/kernel/debug/x86/kaiser-enabled
>>>> 
>>>> and to re-enable it, you can:
>>>> 
>>>>   echo 1 > /sys/kernel/debug/x86/kaiser-enabled
>>>> 
>>>> This is a *minimal* implementation.  There are certainly plenty of
>>>> optimizations that can be done on top of this by using ALTERNATIVES
>>>> among other things.
>>> 
>>> It's not only minimal. It's naive and broken. That thing explodes when
>>> toggled in the wrong moment. I did not even attempt to debug that, because
>>> I think the approach is wrong.
>>> 
>>> If you really want to make it runtime switchable, then:
>>> 
>>> - the shadow tables need to be updated unconditionally. I did not check
>>>  whether thats done right now, but explosions are simpler to achieve when
>>>  switching it back on. Though switching it off crashes as well.
>>> 
>>> - you need to make sure that no task is in user space or on the way to it.
>>>  The much I hate stop_machine(), that's probably the right tool.
>>>  Once everything is in stomp_machine() the switch can be flipped.
>>> 
>>> - the poisoning/unpoisoning of the kernel tables does not need to be done
>>>  from stop_machine(). That can be done from regular context with a TIF
>>>  flag, so you can make sure that every task is up to date before
>>>  returning to user space. Though that needs a lot of thought.
>>> 
>>> For now I really want to see that removed entirely and replaced by a simple
>>> boot time switch. We can use the global variable for now and optimize it
>>> later on.
>>> 
>> 
>> Nah, let's do it right: use either X86_FEATURE_WHATEVER or a
>> static_branch.  We have nice asm support for both.
> 
> Agreed.
> 
>> Keep in mind that, for a static_branch, actually setting the thing needs
>> to be deferred, but that's straightforward.
> 
> That's not an issue during boot. That would be an issue for a run time
> switch.

What I mean is: if you modify a static_branch too early, it blows up terribly.