mm, THP, swap: check whether THP can be split firstly
authorHuang Ying <ying.huang@intel.com>
Thu, 6 Jul 2017 22:37:28 +0000 (15:37 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 6 Jul 2017 23:24:31 +0000 (16:24 -0700)
To swap out THP (Transparent Huage Page), before splitting the THP, the
swap cluster will be allocated and the THP will be added into the swap
cache.  But it is possible that the THP cannot be split, so that we must
delete the THP from the swap cache and free the swap cluster.  To avoid
that, in this patch, whether the THP can be split is checked firstly.
The check can only be done racy, but it is good enough for most cases.

With the patch, the swap out throughput improves 3.6% (from about
4.16GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes.  The test is done on a Xeon E5 v3 system.  The swap
device used is a RAM simulated PMEM (persistent memory) device.  To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.

Link: http://lkml.kernel.org/r/20170515112522.32457-5-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for can_split_huge_page()]
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/huge_mm.h
mm/huge_memory.c
mm/vmscan.c

index a3762d49ba397244f105bade8ee058879c2f08e4..d3b3e8fcc717d517249d73cd587541f6acee1548 100644 (file)
@@ -113,6 +113,7 @@ extern unsigned long thp_get_unmapped_area(struct file *filp,
 extern void prep_transhuge_page(struct page *page);
 extern void free_transhuge_page(struct page *page);
 
+bool can_split_huge_page(struct page *page, int *pextra_pins);
 int split_huge_page_to_list(struct page *page, struct list_head *list);
 static inline int split_huge_page(struct page *page)
 {
@@ -231,6 +232,12 @@ static inline void prep_transhuge_page(struct page *page) {}
 
 #define thp_get_unmapped_area  NULL
 
+static inline bool
+can_split_huge_page(struct page *page, int *pextra_pins)
+{
+       BUILD_BUG();
+       return false;
+}
 static inline int
 split_huge_page_to_list(struct page *page, struct list_head *list)
 {
index 1a168e4bac4bfa7f13aeafaa6ab726fb23e7da72..86975dec0ba160feadfb8aa0d13b8f2be943638d 100644 (file)
@@ -2390,6 +2390,21 @@ int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
        return ret;
 }
 
+/* Racy check whether the huge page can be split */
+bool can_split_huge_page(struct page *page, int *pextra_pins)
+{
+       int extra_pins;
+
+       /* Additional pins from radix tree */
+       if (PageAnon(page))
+               extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
+       else
+               extra_pins = HPAGE_PMD_NR;
+       if (pextra_pins)
+               *pextra_pins = extra_pins;
+       return total_mapcount(page) == page_count(page) - extra_pins - 1;
+}
+
 /*
  * This function splits huge page into normal pages. @page can point to any
  * subpage of huge page to split. Split doesn't change the position of @page.
@@ -2437,7 +2452,6 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
                        ret = -EBUSY;
                        goto out;
                }
-               extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
                mapping = NULL;
                anon_vma_lock_write(anon_vma);
        } else {
@@ -2449,8 +2463,6 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
                        goto out;
                }
 
-               /* Addidional pins from radix tree */
-               extra_pins = HPAGE_PMD_NR;
                anon_vma = NULL;
                i_mmap_lock_read(mapping);
        }
@@ -2459,7 +2471,7 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
         * Racy check if we can split the page, before freeze_page() will
         * split PMDs
         */
-       if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
+       if (!can_split_huge_page(head, &extra_pins)) {
                ret = -EBUSY;
                goto out_unlock;
        }
index 729e37f02de67a171f579add5263c23340825886..abaf0215a35230c6dd2cb7a0fc31a3bf02ea16d8 100644 (file)
@@ -1125,6 +1125,10 @@ static unsigned long shrink_page_list(struct list_head *page_list,
                    !PageSwapCache(page)) {
                        if (!(sc->gfp_mask & __GFP_IO))
                                goto keep_locked;
+                       /* cannot split THP, skip it */
+                       if (PageTransHuge(page) &&
+                           !can_split_huge_page(page, NULL))
+                               goto activate_locked;
                        if (!add_to_swap(page)) {
                                if (!PageTransHuge(page))
                                        goto activate_locked;