Re: [patch 3/5]thp: add tlb_remove_pmd_tlb_entry

[Andrea Arcangeli <aarcange@redhat.com>]

On Fri, Nov 11, 2011 at 02:36:29PM +0800, Shaohua Li wrote:
> On Thu, 2011-11-10 at 23:36 +0800, Andrea Arcangeli wrote:
> > On Tue, Oct 25, 2011 at 10:59:31AM +0800, Shaohua Li wrote:
> > > Index: linux/arch/x86/include/asm/tlb.h
> > > ===================================================================
> > > --- linux.orig/arch/x86/include/asm/tlb.h	2011-10-25 09:00:39.000000000 +0800
> > > +++ linux/arch/x86/include/asm/tlb.h	2011-10-25 09:02:52.000000000 +0800
> > > @@ -4,6 +4,7 @@
> > >  #define tlb_start_vma(tlb, vma) do { } while (0)
> > >  #define tlb_end_vma(tlb, vma) do { } while (0)
> > >  #define __tlb_remove_tlb_entry(tlb, ptep, address) do { } while (0)
> > > +#define __tlb_remove_pmd_tlb_entry(tlb, pmdp, address) do { } while (0)
> > >  #define tlb_flush(tlb) flush_tlb_mm((tlb)->mm)
> > 
> > This is superfluous, it's already define below as noop.
> > 
> > >  
> > >  #include <asm-generic/tlb.h>
> > > Index: linux/include/asm-generic/tlb.h
> > > ===================================================================
> > > --- linux.orig/include/asm-generic/tlb.h	2011-10-25 09:00:23.000000000 +0800
> > > +++ linux/include/asm-generic/tlb.h	2011-10-25 09:18:01.000000000 +0800
> > > @@ -139,6 +139,16 @@ static inline void tlb_remove_page(struc
> > >  		__tlb_remove_tlb_entry(tlb, ptep, address);	\
> > >  	} while (0)
> > >  
> > > +#ifndef __tlb_remove_pmd_tlb_entry
> > > +#define __tlb_remove_pmd_tlb_entry(tlb, pmdp, address) do {} while(0)
> > > +#endif
> > > +
> > > +#define tlb_remove_pmd_tlb_entry(tlb, pmdp, address)		\
> > > +	do {							\
> > > +		tlb->need_flush = 1;				\
> > > +		__tlb_remove_pmd_tlb_entry(tlb, pmdp, address);	\
> > > +	} while (0)
> > 
> > this looks weird, why do we set need_flush = 1 again, considering that
> > we're doing tlb_remove_page() just a few lines later (which also sets
> > tlb->need_flush = 1).
> > 
> > Ok that other archs may need the __tlb_remove_pmd_tlb_entry to be
> > called (and I've no idea why), but the need_flush = 1 seems
> > unnecessary.
> > 
> > Why other archs need the __tlb_remove_pmd_tlb_entry to be called?
> > 
> > One way to go would be to change the tlb->need_flush = 1 in
> > __tlb_remove_page to a VM_BUG_ON(!tlb->need_flush) and then we keep it
> > above and we add the __tlb_remove_pmd_tlb_entry call.
> > 
> > Or is there any place where __tlb_remove_page is called without a
> > tlb_remove_*tlb_entry being called before it?
> > 
> > In any case the VM_BUG_ON will verify this.
> ok, I made the whole tlb_remove_pmd_tlb_entry() noop now. we don't need
> add anything on it for x86 currently. We can change it later if
> necessary.

I thought it'd be cleaner to have only the __tlb_remove_*tlb_entry
variants set need_flush=1 and have __tlb_remove_page just check that
is set under a VM_BUG_ON. That would also avoid a second unnecessary
need_flush = 1 for the pte case which is repeated now (it's not the
repeated in the pmd case in your patch because it's a noop, but the
pte case it's not a noop). Maybe it's not possible but if it's
possible it looks better.