Re: [PATCH v4] mm: fix is_pinnable_page against on cma page

[Jason Gunthorpe]

On Wed, May 11, 2022 at 08:44:43PM -0700, Minchan Kim wrote:
> On Wed, May 11, 2022 at 07:18:56PM -0700, John Hubbard wrote:
> > On 5/11/22 18:08, John Hubbard wrote:
> > > On 5/11/22 18:03, Minchan Kim wrote:
> > > > > 
> > > > > Or there might be some code path that really hates a READ_ONCE() in
> > > > > that place.
> > > > 
> > > > My worry about chaning __get_pfnblock_flags_mask is it's called
> > > > multiple hot places in mm codes so I didn't want to add overhead
> > > > to them.
> > > 
> > > ...unless it really does generate the same code as is already there,
> > > right? Let me check that real quick.
> > > 
> > 
> > It does change the generated code slightly. I don't know if this will
> > affect performance here or not. But just for completeness, here you go:
> > 
> > free_one_page() originally has this (just showing the changed parts):
> > 
> >     mov    0x8(%rdx,%rax,8),%rbx
> >     and    $0x3f,%ecx
> >     shr    %cl,%rbx
> >     and    $0x7,
> > 
> > 
> > And after applying this diff:
> > 
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index 0e42038382c1..df1f8e9a294f 100644
> > +++ b/mm/page_alloc.c
> > @@ -482,7 +482,7 @@ unsigned long __get_pfnblock_flags_mask(const struct
> > page *page,
> >         word_bitidx = bitidx / BITS_PER_LONG;
> >         bitidx &= (BITS_PER_LONG-1);
> > 
> > -       word = bitmap[word_bitidx];
> > +       word = READ_ONCE(bitmap[word_bitidx]);
> >         return (word >> bitidx) & mask;
> >  }
> > 
> > 
> > ...it now does an extra memory dereference:
> > 
> >     lea    0x8(%rdx,%rax,8),%rax
> >     and    $0x3f,%ecx
> >     mov    (%rax),%rbx
> >     shr    %cl,%rbx
> >     and    $0x7,%ebx

Where is the extra memory reference? 'lea' is not a memory reference,
it is just some maths?

> Thanks for checking, John.
> 
> I don't want to have the READ_ONCE in __get_pfnblock_flags_mask 
> atm even though it's an extra memory dereference for specific
> architecutre and specific compiler unless other callsites *do*
> need it.

If a callpath is called under locking or not under locking then I
would expect to have two call chains clearly marked what their locking
conditions are. ie __get_pfn_block_flags_mask_unlocked() - and
obviously clearly document and check what the locking requirements are
of the locked path.

IMHO putting a READ_ONCE on something that is not a memory load from
shared data is nonsense - if a simple == has a stability risk then so
does the '(word >> bitidx) & mask'.

Jason