Re: [RFCv2 00/10] Linear Address Masking enabling

[Thomas Gleixner <tglx@linutronix.de>]

On Thu, May 12 2022 at 21:39, Thomas Gleixner wrote:
> On Thu, May 12 2022 at 10:22, Dave Hansen wrote:
>> One of the stated reasons for adding LAM hardware is that folks want to
>> use sanitizers outside of debugging environments.  To me, that means
>> that LAM is something that the same binary might run with or without.
>
> On/off yes, but is there an actual use case where such a mechanism would
> at start time dynamically chose the number of bits?

This would need cooperation from the application because it has to tell
the magic facility whether it intends to utilize the large VA space on a
5-level paging system or not.

I have no idea how that is supposed to work, but what do I know about
magic.

>> It's totally fine with me if the kernel only initially supports LAM_U57.
>>  But, I'd ideally like to make sure that the ABI can support LAM_U57,
>> LAM_U48, AMD's UAI (in whatever form it settles), or other masks.
>
> Sure. No argument here.

Assumed that the acronym of the day, which uses this, has a real benefit
from the larger number of bits, we can support it.

But we are not going to make this a per thread selectable thing.

It's a process wide decision at startup simply because it does no buy
thread A anything to select U57 if thread B selects U48 before thread A
was able to map something into the U48 covered address space. Same issue
the other way round as then B has to fallback to U57 or NONE. That does
not make any sense at all.

I'm all for flexible, but not just because we can. It has to make sense.

Making it process wide and once on startup puts the 'complexity' into
the prctl(), but keeps the runtime overhead as small as it gets:

  - CR3 switching needs just the | mm->lam_cr3_mask

  - untagging one of the uglies Peter and I came up with

Making U48/U57 hardcoded would not buy much.

Thanks,

        tglx