On Wed, Sep 26, 2012 at 03:38:41PM +0800, Fengguang Wu wrote:
> On Wed, Sep 26, 2012 at 02:06:08AM -0400, Naoya Horiguchi wrote:
> > On Wed, Sep 26, 2012 at 12:02:34AM -0400, Naoya Horiguchi wrote:
> > ...
> > > > > + * page is a thp, not a non-huge compound page.
> > > > > + */
> > > > > + else if (PageTransCompound(page) && !PageSlab(page))
> > > > > u |= 1 << KPF_THP;
> > > >
> > > > Good catch!
> > > >
> > > > Will this report THP for the various drivers that do __GFP_COMP
> > > > page allocations?
> > >
> > > I'm afraid it will. I think of checking PageLRU as an alternative,
> > > but it needs compound_head() to report tail pages correctly.
> > > In this context, pages are not pinned or locked, so it's unsafe to
> > > use compound_head() because it can return a dangling pointer.
> > > Maybe it's a thp's/hugetlbfs's (not kpageflags specific) problem,
> > > so going forward with compound_head() expecting that it will be
> > > fixed in the future work can be an option.
> >
> > It seems that compound_trans_head() solves this problem, so I'll
> > simply use it.
>
> Naoya, in fact I didn't quite catch your concerns. Why not just test
>
> PageTransCompound(page) && PageLRU(page)
If we simply check PageLRU, tail pages in thp only show KPF_COMPOUND_TAIL
and we can't distinguish them from tail pages in non-huge compound pages.
Moreover this behavior is not consistent with that of hugetlbfs tail
pages where tail pages also have KPF_HUGE and are distinct from non-huge
compound pages. I show the output of page-types:
offset len flags
...
2d400 1 ___U_lA____Ma_bH______t____________ (thp head)
2d401 1ff ________________T__________________ (thp tail) # no KPF_THP
...
77000 1 ___U_______Ma__H_G_________________ (hugetlbfs head)
77001 1ff ________________TG_________________ (hugetlbfs tail)
...
11fb50 1 _______________H___________________ (compound head)
11fb51 3 ________________T__________________ (compound tail)
...
11fb58 1 _______S_______H___________________ (slab head)
11fb59 7 ________________T__________________ (slab tail)
H: KPF_COMPOUND_HEAD T: KPF_COMPOUND_TAIL
G: KPF_HUGE t: KPF_THP
So I think it's better to set KPF_THP on thp tail pages.
> The whole page flag report thing is inherently racy and it's fine to
> report wrong values due to races. The "__GFP_COMP reported as THP",
> however, should be avoided because it will make consistent wrong
> reporting of page flags.
Yes, I agree with this point.
Thanks,
Naoya
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>