Re: [PATCHv3 6/8] x86/mm: Provide ARCH_GET_UNTAG_MASK and ARCH_ENABLE_TAGGED_ADDR

[Andy Lutomirski]

On Tue, Jun 28, 2022, at 5:53 PM, Kirill A. Shutemov wrote:
> On Tue, Jun 28, 2022 at 04:42:40PM -0700, Andy Lutomirski wrote:
>> On 6/10/22 07:35, Kirill A. Shutemov wrote:
>> 
>> > +	/* Update CR3 to get LAM active */
>> > +	switch_mm(current->mm, current->mm, current);
>> 
>> Can you at least justify this oddity?  When changing an LDT, we use a
>> dedicated mechanism.  Is there a significant benefit to abusing switch_mm
>> for this?
>
> I'm not sure I follow. LAM mode is set in CR3. switch_mm() has to handle
> it anyway to context switch. Why do you consider it abuse?
>
>> 
>> Also, why can't we enable LAM on a multithreaded process?  We can change an
>> LDT, and the code isn't even particularly complicated.
>
> I reworked this in v4[1] and it allows multithreaded processes. Have you
> got that version?
>
> Intel had issue with mail server, but I assumed it didn't affect my
> patchset since I see it in the archive.
>

I didn’t notice it. Not quite sure what the issue was. Could just be incompetence on my part.

I think that’s the right idea, except that I think you shouldn’t use switch_mm for this. Just update the LAM bits directly.   Once you read mm_cpumask, you should be guaranteed (see next paragraph) that, for each CPU that isn’t in the set, if it switches to the new mm, it will notice the new LAM.

I say “should be” because I think smp_wmb() is insufficient. You’re ordering a write with a subsequent read, which needs smp_mb().

> [1] 
> https://lore.kernel.org/all/20220622162230.83474-1-kirill.shutemov@xxxxxxxxxxxxxxx/
>
> -- 
>  Kirill A. Shutemov