lkml.org 
[lkml]   [2010]   [Mar]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [rfc 5/5] mincore: transparent huge page support
On Wed, Mar 24, 2010 at 11:48:58PM +0100, Andrea Arcangeli wrote:
> On Tue, Mar 23, 2010 at 03:35:02PM +0100, Johannes Weiner wrote:
> > +static int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
> > + unsigned long addr, unsigned long end,
> > + unsigned char *vec)
> > +{
> > + int huge = 0;
> > +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> > + spin_lock(&vma->vm_mm->page_table_lock);
> > + if (likely(pmd_trans_huge(*pmd))) {
> > + huge = !pmd_trans_splitting(*pmd);
>
> Under mmap_sem (read or write) a hugepage can't materialize under
> us. So here the pmd_trans_huge can be lockless and run _before_ taking
> the page_table_lock. That's the invariant I used to keep identical
> performance for all fast paths.

Stupid me. I knew that, I just hadn't internalized it enough to do it
right :)

Btw, unless I miss something else, this is the same in follow_page()?

diff --git a/mm/memory.c b/mm/memory.c
index 22ee158..6c26042 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -1301,18 +1301,14 @@ struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
}
if (pmd_trans_huge(*pmd)) {
spin_lock(&mm->page_table_lock);
- if (likely(pmd_trans_huge(*pmd))) {
- if (unlikely(pmd_trans_splitting(*pmd))) {
- spin_unlock(&mm->page_table_lock);
- wait_split_huge_page(vma->anon_vma, pmd);
- } else {
- page = follow_trans_huge_pmd(mm, address,
- pmd, flags);
- spin_unlock(&mm->page_table_lock);
- goto out;
- }
- } else
+ if (unlikely(pmd_trans_splitting(*pmd))) {
spin_unlock(&mm->page_table_lock);
+ wait_split_huge_page(vma->anon_vma, pmd);
+ } else {
+ page = follow_trans_huge_pmd(mm, address, pmd, flags);
+ spin_unlock(&mm->page_table_lock);
+ goto out;
+ }
/* fall through */
}
if (unlikely(pmd_bad(*pmd)))
> And if it wasn't the case it wouldn't be safe to return huge = 0 as
> the page_table_lock is released at that point.

True.

> > + spin_unlock(&vma->vm_mm->page_table_lock);
> > + /*
> > + * If we have an intact huge pmd entry, all pages in
> > + * the range are present in the mincore() sense of
> > + * things.
> > + *
> > + * But if the entry is currently being split into
> > + * normal page mappings, wait for it to finish and
> > + * signal the fallback to ptes.
> > + */
> > + if (huge)
> > + memset(vec, 1, (end - addr) >> PAGE_SHIFT);
> > + else
> > + wait_split_huge_page(vma->anon_vma, pmd);
> > + } else
> > + spin_unlock(&vma->vm_mm->page_table_lock);
> > +#endif
> > + return huge;
> > +}
> > +
>
> It's probably cleaner to move the block into huge_memory.c and create
> a dummy for the #ifndef version like I did for all the rest.

Agreed.

> I'll incorporate and take care of those changes myself if you don't
> mind, as I'm going to do a new submit for -mm.

Knock yourself out :-)

Hannes


\
 
 \ /
  Last update: 2010-03-25 01:11    [from the cache]
©2003-2011 Jasper Spaans