lkml.org 
[lkml]   [2011]   [Jun]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH] mm: compaction: Abort compaction if too many pages are isolated and caller is asynchronous
On Thu, Jun 02, 2011 at 03:29:54PM +0200, Andrea Arcangeli wrote:
> On Thu, Jun 02, 2011 at 02:03:52AM +0100, Mel Gorman wrote:
> > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > index 2d29c9a..65fa251 100644
> > --- a/mm/huge_memory.c
> > +++ b/mm/huge_memory.c
> > @@ -631,12 +631,14 @@ static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
> > entry = mk_pmd(page, vma->vm_page_prot);
> > entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
> > entry = pmd_mkhuge(entry);
> > +
> > /*
> > - * The spinlocking to take the lru_lock inside
> > - * page_add_new_anon_rmap() acts as a full memory
> > - * barrier to be sure clear_huge_page writes become
> > - * visible after the set_pmd_at() write.
> > + * Need a write barrier to ensure the writes from
> > + * clear_huge_page become visible before the
> > + * set_pmd_at
> > */
> > + smp_wmb();
> > +
>
> On x86 at least this is noop because of the
> spin_lock(&page_table_lock) after clear_huge_page. But I'm not against
> adding this in case other archs supports THP later.
>

I thought spin lock acquisition was one-way where loads/stores
preceeding the lock are allowed to leak into the protected region
but not the other way around?

So we have

clear_huge_page()
__SetPageUptodate(page);
spin_lock(&mm->page_table_lock);
...
set_pmd_at(mm, haddr, pmd, entry);

This spinlock itself does not guarantee that writes from
clear_huge_page are complete before that set_pmd_at().

Whether this is right or wrong, why is the same not true in
collapse_huge_page()? There we are

__collapse_huge_page_copy(pte, new_page, vma, address, ptl);
....
smp_wmb();
spin_lock(&mm->page_table_lock);
...
set_pmd_at(mm, address, pmd, _pmd);

with the comment stressing that this is necessary.

> But smp_wmb() is optimized away at build time by cpp so this can't
> possibly help if you're reproducing !SMP.
>

On X86 !SMP, this is still a barrier() which on gcc is

#define barrier() __asm__ __volatile__("": : :"memory")

so it's a compiler barrier. I'm not working on this at this at the
moment but when I get to it, I'll compare the object files and see
if there are relevant differences. Could be tomorrow before I get
the chance again.

> > page_add_new_anon_rmap(page, vma, haddr);
> > set_pmd_at(mm, haddr, pmd, entry);
> > prepare_pmd_huge_pte(pgtable, mm);
> > @@ -753,6 +755,13 @@ int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
> >
> > pmdp_set_wrprotect(src_mm, addr, src_pmd);
> > pmd = pmd_mkold(pmd_wrprotect(pmd));
> > +
> > + /*
> > + * Write barrier to make sure the setup for the PMD is fully visible
> > + * before the set_pmd_at
> > + */
> > + smp_wmb();
> > +
> > set_pmd_at(dst_mm, addr, dst_pmd, pmd);
> > prepare_pmd_huge_pte(pgtable, dst_mm);
>
> This part seems superfluous to me, it's also noop for !SMP.

Other than being a compiler barrier.

> Only wmb()
> would stay. the pmd is perfectly fine to stay in a register, not even
> a compiler barrier is needed, even less a smp serialization.

There is an explanation in here somewhere because as I write this,
the test machine has survived 14 hours under continual stress without
the isolated counters going negative with over 128 million pages
successfully migrated and a million pages failed to migrate due to
direct compaction being called 80,000 times. It's possible it's a
co-incidence but it's some co-incidence!

--
Mel Gorman
SUSE Labs

\
 
 \ /
  Last update: 2011-06-02 16:53    [W:0.115 / U:0.260 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site