Re: [RFT PATCH v2 03/23] x86/boot: Drop global variables keeping track of LA57 state

[Ard Biesheuvel]

On Sun, 4 May 2025 at 16:58, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> On Sun, 4 May 2025 at 06:51, Ingo Molnar <mingo@xxxxxxxxxx> wrote:
> >
> > Cannot pgtable_l5_enabled() be a single, simple percpu flag or so?
>
> Isn't this logically something that should just be a static key? For
> some reason I thought we did that already, but looking around, that
> was only in the critical user access routines (and got replaced by the
> 'runtime-const' stuff)
>

The ordinary, late version of pgtable_l5_enabled() is #define'd to
cpu_feature_enabled(), which is runtime patched in a slightly more
efficient manner than the alternative I'm proposing here. It is
conceptually the same as a static key, i.e., the asm goto() gets
patched into a JMP or a NOP.

Whether or not using runtime constants are worth the hassle for
pgdir_shift and ptrs_per_p4d is a different question, I'll keep them
as ordinary globals for now, and drop the conditional expressions.

> But I guess that what Ard wants to get rid of is the variable itself,
> and for early boot using static keys sounds like a bad idea.
>

I wanted to get rid of the variable, and then noticed the nasty
'#define USE_EARLY_PGTABLE_L5' thing, so I attempted to fix that as
well. (Having to keep track of which code may be called early, late or
both is how I ended up in this swamp to begin with.)

> Honestly, looking at this, I think we should fix the *users* of
> pgtable_l5_enabled().
>
> Because I think there are two distinct classes:
>
>  - the stuff in <asm/pgtable.h> is exposed as the generic page table
> accessor macros to "real" code, and should probably use a static key
> (ie things like pgd_clear() / pgd_none() and friends)
>
>  - but in code like __kernel_physical_mapping_init() feels to me like
> using the value in %cr4 actually makes sense
>
> but that looks like a much bigger and fundamental patch than Ard's.
>

I think we should just rely on cpu_feature_enabled(), which can in
fact be called early as long as the capability gets set. So instead of
setting __pgtable_l5_enabled, we should just call
'set_cpu_cap(&boot_cpu_data, X86_FEATURE_LA57)' so that even early
callers can use the ordinary pgtable_l5_enabled(), and we can drop the
early flavor. (The decompressor will need its own version but it can
just inspect CR4.LA57 directly)