On Fri, Jun 18, 2021 at 10:32:42AM +0530, Aneesh Kumar K.V wrote:
> On 6/17/21 11:32 PM, Nathan Chancellor wrote:
> > Rebuilt the CC list because most people were added based on the
> > incorrect bisect result.
> >
> > On Thu, Jun 17, 2021 at 02:51:49PM +0100, Matthew Wilcox wrote:
> > > On Thu, Jun 17, 2021 at 06:15:45PM +0530, Naresh Kamboju wrote:
> > > > On Thu, 17 Jun 2021 at 17:41, Naresh Kamboju <naresh.kamboju@xxxxxxxxxx> wrote:
> > > > > x86_64-linux-gnu-ld: mm/mremap.o: in function `move_pgt_entry':
> > > > > mremap.c:(.text+0x763): undefined reference to `__compiletime_assert_342'
> > > >
> > > > The git bisect pointed out the first bad commit.
> > > >
> > > > The first bad commit:
> > > > commit 928cf6adc7d60c96eca760c05c1000cda061604e
> > > > Author: Stephen Boyd <swboyd@xxxxxxxxxxxx>
> > > > Date: Thu Jun 17 15:21:35 2021 +1000
> > > > module: add printk formats to add module build ID to stacktraces
> > >
> > > Your git bisect probably went astray. There's no way that commit
> > > caused that regression.
> >
> > My bisect landed on commit 83f85ac75855 ("mm/mremap: convert huge PUD
> > move to separate helper"). flush_pud_tlb_range() evaluates to
> > BUILD_BUG() when CONFIG_TRANSPARENT_HUGEPAGE is unset but this function
> > is present just based on the value of
> > CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD.
> >
> > $ make -skj(nproc) ARCH=x86_64 CC=clang O=build/x86_64 distclean allnoconfig mm/mremap.o
> >
> > $ llvm-readelf -s build/x86_64/mm/mremap.o &| rg __compiletime_assert
> > 21: 0000000000000000 0 NOTYPE GLOBAL DEFAULT UND __compiletime_assert_337
> >
> > $ rg TRANSPARENT_ build/x86_64/.config
> > 450:CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE=y
> > 451:CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD=y
> > 562:# CONFIG_TRANSPARENT_HUGEPAGE is not set
> >
> > Not sure why this does not happen on newer clang versions, presumably
> > something with inlining decisions? Still seems like a legitimate issue
> > to me.
> >
>
> gcc 10 also doesn't give a build error. I guess that is because we evaluate
>
> if (pud_trans_huge(*old_pud) || pud_devmap(*old_pud)) {
>
> to if (0) with CONFIG_TRANSPARENT_HUGEPAGE disabled.
>
> switching that to if (1) do results in BUILD_BUG triggering.
Thanks for pointing that out. I think what happens with clang-10 and
clang-11 is that move_huge_pud() gets inlined into move_pgt_entry() but
then the compiler does not figure out that the HPAGE_PUD case is dead so
the code sticks around, where as GCC and newer clang versions can figure
that out and eliminate that case.
> Should we fix this ?
Yes, I believe that we should.
> modified mm/mremap.c
> @@ -336,7 +336,7 @@ static inline bool move_normal_pud(struct vm_area_struct
> *vma,
> }
> #endif
>
> -#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
> +#if defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD) &&
> defined(CONFIG_TRANSPARENT_HUGEPAGE)
> static bool move_huge_pud(struct vm_area_struct *vma, unsigned long
> old_addr,
> unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
> {
That works or we could mirror what has already been done for the
HPAGE_PMD case. No personal preference.
diff --git a/mm/mremap.c b/mm/mremap.c
index 9a7fbec31dc9..5989d3990020 100644
--- a/mm/mremap.c
+++ b/mm/mremap.c
@@ -460,7 +460,8 @@ static bool move_pgt_entry(enum pgt_entry entry, struct vm_area_struct *vma,
new_entry);
break;
case HPAGE_PUD:
- moved = move_huge_pud(vma, old_addr, new_addr, old_entry,
+ moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
+ move_huge_pud(vma, old_addr, new_addr, old_entry,
new_entry);
break;
Cheers,
Nathan
