On Thu, Jun 04, 2020 at 12:17:23PM +0100, Will Deacon wrote:
> Hi, [+Peter]
>
> On Thu, Jun 04, 2020 at 10:48:03AM +0000, Christophe Leroy wrote:
> > Using mpc885_ads_defconfig with CONFIG_PPC_16K_PAGES instead of
> > CONFIG_PPC_4K_PAGES, getting the following build failure:
> >
> > CC mm/gup.o
> > In file included from ./include/linux/kernel.h:11:0,
> > from mm/gup.c:2:
> > In function 'gup_hugepte.constprop',
> > inlined from 'gup_huge_pd.isra.78' at mm/gup.c:2465:8:
> > ./include/linux/compiler.h:392:38: error: call to '__compiletime_assert_257'
> > declared with attribute error: Unsupported access size for
> > {READ,WRITE}_ONCE().
> > _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__)
> > ^
> > ./include/linux/compiler.h:373:4: note: in definition of macro
> > '__compiletime_assert'
> > prefix ## suffix(); \
> > ^
> > ./include/linux/compiler.h:392:2: note: in expansion of macro
> > '_compiletime_assert'
> > _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__)
> > ^
> > ./include/linux/compiler.h:405:2: note: in expansion of macro
> > 'compiletime_assert'
> > compiletime_assert(__native_word(t) || sizeof(t) == sizeof(long long), \
> > ^
> > ./include/linux/compiler.h:291:2: note: in expansion of macro
> > 'compiletime_assert_rwonce_type'
> > compiletime_assert_rwonce_type(x); \
> > ^
> > mm/gup.c:2428:8: note: in expansion of macro 'READ_ONCE'
> > pte = READ_ONCE(*ptep);
> > ^
> > In function 'gup_get_pte',
> > inlined from 'gup_pte_range' at mm/gup.c:2228:9,
> > inlined from 'gup_pmd_range' at mm/gup.c:2613:15,
> > inlined from 'gup_pud_range' at mm/gup.c:2641:15,
> > inlined from 'gup_p4d_range' at mm/gup.c:2666:15,
> > inlined from 'gup_pgd_range' at mm/gup.c:2694:15,
> > inlined from 'internal_get_user_pages_fast' at mm/gup.c:2785:3:
>
> At first glance, this looks like a real bug in the 16k page code -- you're
> loading the pte non-atomically on the fast GUP path and so you're prone to
> tearing, which probably isn't what you want. For a short-term hack, I'd
> suggest having CONFIG_HAVE_FAST_GUP depend on !CONFIG_PPC_16K_PAGES, but if
> you want to support this them you'll need to rework your pte_t so that it
> can be loaded atomically.
Looking at commit 55c8fc3f49302, they're all the exact same value, so
what they could do is grow another special gup_get_pte() variant that
just loads the first value.
Also, per that very same commit, there's a distinct lack of WRITE_ONCE()
in the pte_update() / __set_pte_at() paths for much of Power.
