On Tue, May 22, 2012 at 10:48 AM, Andrea Arcangeli [off-list ref] wrote:

On Thu, May 17, 2012 at 03:17:49PM -0700, Pravin B Shelar wrote:

quoted

diff --git a/mm/swap.c b/mm/swap.c
index 8ff73d8..44a0f81 100644
--- a/mm/swap.c
+++ b/mm/swap.c

@@ -82,6 +82,19 @@ static void put_compound_page(struct page *page)

              if (likely(page != page_head &&
                         get_page_unless_zero(page_head))) {
                      unsigned long flags;
+
+                     if (PageSlab(page_head)) {
+                             if (PageTail(page)) {
+                                     /* THP can not break up slab pages, avoid
+                                      * taking compound_lock(). */
+                                     if (put_page_testzero(page_head))
+                                             VM_BUG_ON(1);
+
+                                     atomic_dec(&page->_mapcount);
+                                     goto skip_lock_tail;
+                             } else
+                                     goto skip_lock;
+                     }

Some commentary on the fact slab prefers not using atomic ops on the
page->flags could help here.

ok.

quoted

                      /*
                       * page_head wasn't a dangling pointer but it
                       * may not be a head page anymore by the time

@@ -93,6 +106,7 @@ static void put_compound_page(struct page *page)

                              /* __split_huge_page_refcount run before us */
                              compound_unlock_irqrestore(page_head, flags);
                              VM_BUG_ON(PageHead(page_head));

Hmmm hmmm while reviewing this one, I've been thinking maybe the head
page after the hugepage split, could have been freed and reallocated
as order 1 or 2, and legitimately become an head page again.

The whole point of the bug-on is that it cannot be reallocated as a
THP beause the tail is still there and it's not free yet, but it
doesn't take into account the head page could be allocated as a
compound page of a smaller size and maybe the tail is the last subpage
of the thp.

So there's the risk of a false positive, in an extremely unlikely case
(the fact slab goes in unmovable pageblocks and thp goes in movable
further decreases the probability). All production kernels runs with
VM_BUG_ON disabled so it's a very small concern, but maybe we should
delete it. It has never triggered, just code reivew. Do you agree?

right, I will delete it.

quoted

+                     skip_lock:
                              if (put_page_testzero(page_head))
                                      __put_single_page(page_head);
                      out_put_single:

@@ -115,6 +129,8 @@ static void put_compound_page(struct page *page)

                      VM_BUG_ON(atomic_read(&page_head->_count) <= 0);
                      VM_BUG_ON(atomic_read(&page->_count) != 0);
                      compound_unlock_irqrestore(page_head, flags);
+
+                     skip_lock_tail:
                      if (put_page_testzero(page_head)) {
                              if (PageHead(page_head))
                                      __put_compound_page(page_head);

@@ -162,6 +178,15 @@ bool __get_page_tail(struct page *page)

      struct page *page_head = compound_trans_head(page);

      if (likely(page != page_head && get_page_unless_zero(page_head))) {
+
+             if (PageSlab(page_head)) {
+                     if (likely(PageTail(page))) {
+                             __get_page_tail_foll(page, false);
+                             return true;
+                     } else
+                             goto out;
+             }
+

A comment here too would be nice.

quoted

              /*
               * page_head wasn't a dangling pointer but it
               * may not be a head page anymore by the time

@@ -175,6 +200,8 @@ bool __get_page_tail(struct page *page)

                      got = true;
              }
              compound_unlock_irqrestore(page_head, flags);
+
+             out:
              if (unlikely(!got))
                      put_page(page_head);

out could go in the line below. Assuming we don't want to be cleaner
and use put_page above instead of goto, that would also drop a branch
probably (the goto place is such a slow path). I'm fine either ways.

It's not the cleanest of the patches but it's clearly a performance
tweak.

ok, I will post revised patch.

Reviewed-by: Andrea Arcangeli <redacted>

Thanks,
Andrea

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>