Re: [PATCH v2 3/5] riscv: support early isa ext and use it to optimize pgtable_l4|l5_enabled

[Jisheng Zhang <jszhang@kernel.org>]

On Wed, Sep 09, 2026 at 12:49:09PM +0100, Conor Dooley wrote:
> On Wed, Sep 09, 2026 at 07:46:49AM +0800, Jisheng Zhang wrote:
> > > > This assumes SV48 and SV57 are the only "early" users, but
> > > > other ISA ext may also need "early", who knows. So if future some extensions
> > > > need the "early", the code is ready, the author doesn't need to care about
> > > > the isa filling at all.
> > > > 
> > > > So I prefer my patch as is. What's your opinion?
> > > 
> > > We're about decade into the port being merged and this is the only thing
> > 
> > but the isa extension alternative is not, it was introduced by me three
> > years ago. But no matter how many years, 
> > 
> > > behaving in this way. I don't think the code in this patch is worth it on
> > > the off chance that something else comes along. If it does, we can
> > 
> > this makes sense. Let me cook a new version
> > 
> > > always fish this implementation back up and use it.
> > 
> > > 
> > > I'd also like to differentiate this code from "needing early", because
> > > this is about populating the information early in the extension bitmap,
> > > rather than about actually needing the information. There's no advantage
> > 
> > the code after arch/riscv/mm/init.c but before mmu on needs the
> > bitmap informaion when pgtable_l5|l4_enabled() is called (w/o
> > USE_EARLY_PGTABLE_LEVELS)
> 
> Oh, I must have missed something then, I didn't realise you were
> patching the alternatives early - I thought you were skipping them until
> the information became available at the normal time. Sorry bout that.


Aha, your suggestion about the simplifying the isa bitmap filling
still works.

> 
> How bad is the damage btw, if you implement pgtable_l5_enabled() as
> 
> static __always_inline bool pgtable_l5_enabled(void)
> {
> 	if (riscv_has_extension_likely(RISCV_ISA_EXT_SV57))
> 		return true;
> 
> 	return _pgtable_l5_enabled;
> }
> 
> ? It shouldn't be too bad since it should get expanded to stuff like
> 
> 	if (riscv_has_extension_likely(RISCV_ISA_EXT_SV57) || _pgtable_l5_enabled)
> 
> by the compiler. I wonder if you can do this and get rid of
> USE_EARLY_PGTABLE_LEVELS entirely, since the unpatched alternative
> should return false?

I know why you want this. AIUI, the early patched alternative is
still needed outside arch/riscv/mm/init.c but before mmu on. For
example, on a SV57 capable platform, the set_satp_mode() will detect
pgtable cap correctly, then _pgtable_l5_enabled = true and
_pgtable_l4_enabled = true, then any pgtable_l5|l4_enabled() calling
outside init.c but before mmu on also expects true. so we need to set
the isa bitmap so that latter apply_early_boot_alternatives() can
correctly patch .text section for us.

BTW, the USE_EARLY_PGTABLE_LEVELS is stolen from the x86 world, see
its USE_EARLY_PGTABLE_L5 ;)

> 
> Cheers,
> Conor.
> 
> (btw, I am kinda unavailable til the 22nd, so sorry if I take some time
> to reply here or to a new revision)

I just sent out v3, could you plz kindly review it when convenient

> 
> > 
> > > gained, as far as I can tell, by setting this early and it only makes
> > > the code more complicated.
> > > 
> >