On Tue, 2018-10-23 at 10:30 -0700, Mike Kravetz wrote:
..... snip.... Here is updated patch without the drop_caches change and updated fixes tag.
From: Mike Kravetz mike.kravetz@oracle.com
hugetlbfs: dirty pages as they are added to pagecache
Some test systems were experiencing negative huge page reserve counts and incorrect file block counts. This was traced to /proc/sys/vm/drop_caches removing clean pages from hugetlbfs file pagecaches. When non-hugetlbfs explicit code removes the pages, the appropriate accounting is not performed.
This can be recreated as follows: fallocate -l 2M /dev/hugepages/foo echo 1 > /proc/sys/vm/drop_caches fallocate -l 2M /dev/hugepages/foo grep -i huge /proc/meminfo AnonHugePages: 0 kB ShmemHugePages: 0 kB HugePages_Total: 2048 HugePages_Free: 2047 HugePages_Rsvd: 18446744073709551615 HugePages_Surp: 0 Hugepagesize: 2048 kB Hugetlb: 4194304 kB ls -lsh /dev/hugepages/foo 4.0M -rw-r--r--. 1 root root 2.0M Oct 17 20:05 /dev/hugepages/foo
To address this issue, dirty pages as they are added to pagecache. This can easily be reproduced with fallocate as shown above. Read faulted pages will eventually end up being marked dirty. But there is a window where they are clean and could be impacted by code such as drop_caches. So, just dirty them all as they are added to the pagecache.
Fixes: 6bda666a03f0 ("hugepages: fold find_or_alloc_pages into huge_no_page()") Cc: stable@vger.kernel.org Signed-off-by: Mike Kravetz mike.kravetz@oracle.com
mm/hugetlb.c | 6 ++++++ 1 file changed, 6 insertions(+)
diff --git a/mm/hugetlb.c b/mm/hugetlb.c index 5c390f5a5207..7b5c0ad9a6bd 100644 --- a/mm/hugetlb.c +++ b/mm/hugetlb.c @@ -3690,6 +3690,12 @@ int huge_add_to_page_cache(struct page *page, struct address_space *mapping, return err; ClearPagePrivate(page);
- /*
* set page dirty so that it will not be removed fromcache/file
* by non-hugetlbfs specific code paths.*/- set_page_dirty(page);
- spin_lock(&inode->i_lock); inode->i_blocks += blocks_per_huge_page(h); spin_unlock(&inode->i_lock);
This looks good.
Reviewed-by: Khalid Aziz khalid.aziz@oracle.com
-- Khalid