On Thu, May 23, 2024 at 08:30:19PM -0700, Dave Hansen wrote: > On 5/23/24 16:07, Peter Xu wrote: > > Probably not.. I think I can define a pgprot_to_large() globally, pointing > > that to pgprot_4k_2_large() on x86 and make the fallback to be noop. And > > if there's a new version I'll guarantee to run over my cross compilers. > > I guess that would be functional, but it would be a bit mean to > everybody else. > > > Any comments on the idea itself? Do we have a problem, or maybe I > > overlooked something? > > I think it's probably unnecessary to inflict this particular x86-ism on > generic code. The arch-generic 'prot' should have PAT at its 4k > (_PAGE_BIT_PAT) position and then p*d_mkhuge() can shift it into the > _PAGE_BIT_PAT_LARGE spot. Right that's another option indeed. It's just that I found it might in many cases be better when we have the API separately properly and making the pairs matching each other. For example, it could be clearer if pxx_mkhuge() does exactly what pxx_leaf() would check against. PS: I hoped it's called pxx_huge() already to make the name paired with each other; afaict we called it pxx_leaf() only because pxx_huge() used to be "abused" by hugetlbfs before.. now it's gone. The other thing is we mostly only need these knobs for special maps like pfnmaps, am I right? OTOH we use WB for RAMs, and maybe we don't want to bother any PAT stuff when the kernel is installing a THP anonymous? IMHO having pgprot_to_large() is fine even if only x86 has it; it's really like pfn tracking itself which is noop for !x86. but I'll follow your advise if you still insist; I don't really have a strong opinion. But if so I'd also like to mention a 3rd option, which is to have pxx_mkhuge_prot(), fallback to pxx_mkhuge() for !x86. That'll make pxx_huge() untainted for x86. I'm not sure whether that would ease the same concern, though. In all cases, thanks for confirming this issue, I appreciate that. Let me know if you have any comment on patch 1 too; that one isn't a problem so far iiuc, but it can be soon. Thanks, -- Peter Xu
