Re: [PATCHv2 5/8] khugepaged: Allow to callapse a page shared across fork

From: Yang Shi
Date: Thu Apr 09 2020 - 20:03:09 EST




On 4/8/20 11:51 AM, Yang Shi wrote:


On 4/8/20 6:10 AM, Kirill A. Shutemov wrote:
On Mon, Apr 06, 2020 at 01:50:56PM -0700, Yang Shi wrote:

On 4/3/20 4:29 AM, Kirill A. Shutemov wrote:
The page can be included into collapse as long as it doesn't have extra
pins (from GUP or otherwise).

Signed-off-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
---
ÂÂ mm/khugepaged.c | 25 ++++++++++++++-----------
ÂÂ 1 file changed, 14 insertions(+), 11 deletions(-)

diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index 57ff287caf6b..1e7e6543ebca 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -581,11 +581,18 @@ static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
ÂÂÂÂÂÂÂÂÂÂ }
ÂÂÂÂÂÂÂÂÂÂ /*
-ÂÂÂÂÂÂÂÂ * cannot use mapcount: can't collapse if there's a gup pin.
-ÂÂÂÂÂÂÂÂ * The page must only be referenced by the scanned process
-ÂÂÂÂÂÂÂÂ * and page swap cache.
+ÂÂÂÂÂÂÂÂ * Check if the page has any GUP (or other external) pins.
+ÂÂÂÂÂÂÂÂ *
+ÂÂÂÂÂÂÂÂ * The page table that maps the page has been already unlinked
+ÂÂÂÂÂÂÂÂ * from the page table tree and this process cannot get
+ÂÂÂÂÂÂÂÂ * additinal pin on the page.
+ÂÂÂÂÂÂÂÂ *
+ÂÂÂÂÂÂÂÂ * New pins can come later if the page is shared across fork,
+ÂÂÂÂÂÂÂÂ * but not for the this process. It is fine. The other process
+ÂÂÂÂÂÂÂÂ * cannot write to the page, only trigger CoW.
ÂÂÂÂÂÂÂÂÂÂÂ */
-ÂÂÂÂÂÂÂ if (page_count(page) != 1 + PageSwapCache(page)) {
+ÂÂÂÂÂÂÂ if (total_mapcount(page) + PageSwapCache(page) !=
+ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ page_count(page)) {
This check looks fine for base page, but what if the page is PTE-mapped THP?
The following patch made this possible.

If it is PTE-mapped THP and the page is in swap cache, the refcount would be
512 + the number of PTE-mapped pages.

Shall we do the below change in the following patch?

extra_pins = PageSwapCache(page) ? nr_ccompound(page) - 1 : 0;
if (total_mapcount(page) + PageSwapCache(page) != page_count(page) -
extra_pins) {
...
Looks like you're right.

It would be nice to have a test case to demonstrate the issue.

Is there any way to trigger moving the page to swap cache? I don't see it
immediately.

It sounds not easy to trigger since it totally depends on timing, I'm wondering we may have to use MADV_PAGEOUT? Something below off the top of my head may trigger this?


ÂÂÂ CPUÂÂÂ ÂÂ AÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ Â ÂÂÂ Â ÂÂÂ ÂÂÂ CPUÂÂÂ B ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ CPUÂÂÂ C
In parent:
MADV_HUGEPAGE
page fault to fill with THP
fork
ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ In Child:
MADV_NOHUGEPAGE
MADV_DONTNEED (split pmd)
MADV_PAGEOUT
ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ -> add_to_swap
ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ khugepaged scan parent and try to collapse PTE-mapped

ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ ÂÂÂ -> try_to_unmap

When doing MADV_DONTNEED we need make sure head page is unmapped since MADV_PAGEOUT would call page_mapcount(page) to skip shared mapping.

This can't trigger it since MADV_PAGEOUT would isolate the page from lru so khugepaged_scan_pmd() can't reach the refcount check at all.

If try_to_unmap() or pageout() fails, the page may stay in swap cache. But I didn't think of an easy way to make this happen.