Re: [PATCH v2] x86: mm: Do not use set_{pud,pmd}_safe when splitting the large page

[Peter Zijlstra]

On Mon, Apr 15, 2019 at 08:58:52AM -0700, Dave Hansen wrote:
> On 4/15/19 7:55 AM, Singh, Brijesh wrote:
> >  static unsigned long __meminit
> >  phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
> > -	      pgprot_t prot)
> > +	      pgprot_t prot, bool safe)
> >  {
> >  	unsigned long pages = 0, paddr_next;
> >  	unsigned long paddr_last = paddr_end;
> > @@ -432,7 +463,7 @@ phys_pte_init(pte_t *pte_page, unsigned long paddr, unsigned long paddr_end,
> >  					     E820_TYPE_RAM) &&
> >  			    !e820__mapped_any(paddr & PAGE_MASK, paddr_next,
> >  					     E820_TYPE_RESERVED_KERN))
> > -				set_pte_safe(pte, __pte(0));
> > +				__set_pte(pte, __pte(0), safe);
> >  			continue;
> >  		}
> 
> The changelog is great, btw.
> 
> But, I'm not a big fan of propagating the 'safe' nomenclature.  Could
> we, at least, call it 'overwrite_safe' or something if we're going to
> have a variable name?  Or even, 'new_entries_only' or something that
> actually conveys meaning?
> 
> Because, just reading it, I always wonder "why do we have an unsafe
> variant, that's stupid" every time. :)

s/safe/init/ on the whole thing?

And maybe even back on the initial _safe functions? Because all of this
is about initializing page-tables, which is a TLB *safe* operation I
suppose :-)

> > +#define DEFINE_ENTRY(type1, type2, safe)			\
> > +static inline void __set_##type1(type1##_t *arg1,		\
> > +			type2##_t arg2, bool safe)		\
> > +{								\
> > +	if (safe)						\
> > +		set_##type1##_safe(arg1, arg2);			\
> > +	else							\
> > +		set_##type1(arg1, arg2);			\
> > +}
> 
> While I appreciate the brevity that these macros allow, I detest their
> ability to thwart cscope and grep.  I guess it's just one file, but it
> does make me grumble a bit.

There is scripts/tags.sh where you can add to regex_c to teach
cscope/ctags about magic macros.

> Also, can we do better than "__"?  Aren't these specific to
> initialization, and only for the kernel?  Maybe we should call them
> meminit_set_pte() or kern_set_pte() or something so make it totally
> clear to the reader that they're new.

set_*_init() and set_*() I suppose.

> 
> > -		kernel_physical_mapping_init(__pa(vaddr & pmask),
> > -					     __pa((vaddr_end & pmask) + psize),
> > -					     split_page_size_mask);
> > +		kernel_physical_mapping_change(__pa(vaddr & pmask),
> > +					       __pa((vaddr_end & pmask) + psize),
> > +					       split_page_size_mask);
> 
> BTW, this hunk is really nice the way that the new naming makes it more
> intuitive what's going on.  My only nit w9uld be that we now have two
> very similarly-named functions with different TLB-flushing requirements.
> 
> Could we please include a comment at this site that reminds us that we
> owe a TLB flush after this?

Right, a comment would be good. I think my initial proposal had the TLB
flushing inside, but I see the usage is in a loop, so I appreciate the
desire to keep the TLB flushing outside.