Re: [PATCH] mm,thp: recheck each page before collapsing file THP

[Song Liu]

> On Oct 18, 2019, at 8:41 AM, Rik van Riel <riel@xxxxxx> wrote:
> 
> On Fri, 2019-10-18 at 16:34 +0300, Kirill A. Shutemov wrote:
>> On Thu, Oct 17, 2019 at 10:08:32PM -0700, Song Liu wrote:
>>> In collapse_file(), after locking the page, it is necessary to
>>> recheck
>>> that the page is up-to-date, clean, and pointing to the proper
>>> mapping.
>>> If any check fails, abort the collapse.
>>> 
>>> Fixes: 99cb0dbd47a1 ("mm,thp: add read-only THP support for (non-
>>> shmem) FS")
>>> Cc: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
>>> Cc: Johannes Weiner <hannes@xxxxxxxxxxx>
>>> Cc: Hugh Dickins <hughd@xxxxxxxxxx>
>>> Cc: William Kucharski <william.kucharski@xxxxxxxxxx>
>>> Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
>>> Signed-off-by: Song Liu <songliubraving@xxxxxx>
>>> ---
>>> mm/khugepaged.c | 8 ++++++++
>>> 1 file changed, 8 insertions(+)
>>> 
>>> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
>>> index 0a1b4b484ac5..7da49b643c4d 100644
>>> --- a/mm/khugepaged.c
>>> +++ b/mm/khugepaged.c
>>> @@ -1619,6 +1619,14 @@ static void collapse_file(struct mm_struct
>>> *mm,
>>> 				result = SCAN_PAGE_LOCK;
>>> 				goto xa_locked;
>>> 			}
>>> +
>>> +			/* double check the page is correct and clean
>>> */
>>> +			if (unlikely(!PageUptodate(page)) ||
>>> +			    unlikely(PageDirty(page)) ||
>>> +			    unlikely(page->mapping != mapping)) {
>>> +				result = SCAN_FAIL;
>>> +				goto out_unlock;
>>> +			}
>>> 		}
>>> 
>>> 		/*
>> 
>> Hm. But why only for !is_shmem? Or I read it wrong?
> 
> It looks like the shmem code path has its own way of bailing
> out when a page is !PageUptodate. Also, shmem can handle dirty
> pages fine.

Seems the PageUptodate check is still necessary for shmem? 
shmem_getpage() makes sure the page is uptodate, but these is still
a small window that the page could become !uptodate. 

> 
> However, I suppose the shmem code might want to check for truncated
> pages, which it does not curretnly appear to do. I guess doing
> the trylock_page under the xarray lock may protect against truncate,
> but that is subtle enough that at the very least it should be
> documented.

Johannes pointed out in our internal code review that, there is 
already a page_mapping() check later in the function. 

PageDirty check is only necessary for !is_shmem. And it should not 
happen, because we only support read-only text. Adding a warning for
it. Also, move PageDirty check to after page_mapping() check, because
if truncate happens, the PageDirty doesn't violent the read-only
assumption. 

Overall, I guess we need something like:

============================= 8< =============================
diff --git c/mm/khugepaged.c w/mm/khugepaged.c
index 0a1b4b484ac5..40c549302d36 100644
--- c/mm/khugepaged.c
+++ w/mm/khugepaged.c
@@ -1626,7 +1626,12 @@ static void collapse_file(struct mm_struct *mm,
                 * without racing with truncate.
                 */
                VM_BUG_ON_PAGE(!PageLocked(page), page);
-               VM_BUG_ON_PAGE(!PageUptodate(page), page);
+
+               /* double check the page is up to date */
+               if (unlikely(!PageUptodate(page))) {
+                       result = SCAN_FAIL;
+                       goto out_unlock;
+               }

                /*
                 * If file was truncated then extended, or hole-punched, before
@@ -1642,6 +1647,15 @@ static void collapse_file(struct mm_struct *mm,
                        goto out_unlock;
                }

+               /*
+                * khugepaged should not try to collapse dirty pages for
+                * file THP. Show warning if this somehow happens.
+                */
+               if (WARN_ON_ONCE(!is_shmem && PageDirty(page))) {
+                       result = SCAN_FAIL;
+                       goto out_unlock;
+               }
+
                if (isolate_lru_page(page)) {
                        result = SCAN_DEL_PAGE_LRU;
                        goto out_unlock;

============================= 8< =============================

Thanks,
Song