Re: [PATCHv2 0/4] x86: 5-level related changes into decompression code

[H. Peter Anvin]

On 11/29/17 14:31, Borislav Petkov wrote:
> 
> A couple of points:
> 
> * so this box here has a normal grub installation and apparently grub
> jumps to some other entry point.
> 

Yes, Grub as a matter of policy(!) does everything in the most braindead
way possible.  You have to use "linux16" or "linuxefi" to make it do
something sane.

> * I'm not convinced we need to do everything you typed because this is
> only a temporary issue and once X86_5LEVEL is complete, it should work.
> I mean, it needs to work otherwise forget single-system image and I
> don't think we want to give that up.
> 
>> However, if the bootloader jumps straight into the code what do you
>> expect it to do?  We have no real concept about what we'd need to do to
>> issue a message as we really don't know what devices are available on
>> the system, etc.  If the screen_info field in struct boot_params has
>> been initialized then we actually *do* know how to write to the screen
>> -- if you are okay with including a text font etc. since modern systems
>> boot in graphics mode.
> 
> We switch to text mode and dump our message. Can we do that?

What is text mode?  It is hardware that is going away(*), and you don't
even know if you have a display screen on your system at all, or how
you'd have to configure your display hardware even if it is "mostly" VGA.

> I wouldn't want to do any of this back'n'forth between kernel and boot
> loader because that sounds fragile, at least to me. And again, I'm
> not convinced we should spend too much energy on this as the issue is
> temporary AFAICT.

Well, it's not just limited to 5-level mode; it's kind a general issue.
We have had this issue for a very, very long time -- all the way back to
i386 PAE at the very least.  I'm personally OK with triple-faulting the
CPU in this case.

	-hpa


(*) And for good reason -- it is completely memory-latency-bound as you
    have an indirect reference for every byte you fetch.  In a UMA
    system this sucks up an insane amount of system bandwidth, unless
    you are willing to burn the area of having a 16K SRAM cache.

    VGA hardware, additionally, has a bunch of insane operations that
    have to be memory-mapped.  The resulting hardware screws with
    pretty much any sane GPU implementation, so I'm fully expecting that
    as soon as GPUs no longer come with a CBIOS option ROM VGA hardware
    will be dropped more or less immediately.