ext4: bugfix for mmaped pages in mpage_release_unused_pages()
authorwangguang <wang.guang55@zte.com.cn>
Thu, 15 Sep 2016 15:32:46 +0000 (11:32 -0400)
committerTheodore Ts'o <tytso@mit.edu>
Thu, 15 Sep 2016 15:32:46 +0000 (11:32 -0400)
Pages clear buffers after ext4 delayed block allocation failed,
However, it does not clean its pte_dirty flag.
if the pages unmap ,in cording to the pte_dirty ,
unmap_page_range may try to call __set_page_dirty,

which may lead to the bugon at
mpage_prepare_extent_to_map:head = page_buffers(page);.

This patch just call clear_page_dirty_for_io to clean pte_dirty
at mpage_release_unused_pages for pages mmaped.

Steps to reproduce the bug:

(1) mmap a file in ext4
addr = (char *)mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_SHARED,
                    fd, 0);
memset(addr, 'i', 4096);

(2) return EIO at

ext4_writepages->mpage_map_and_submit_extent->mpage_map_one_extent

which causes this log message to be print:

                ext4_msg(sb, KERN_CRIT,
                        "Delayed block allocation failed for "
                        "inode %lu at logical offset %llu with"
                        " max blocks %u with error %d",
                        inode->i_ino,
                        (unsigned long long)map->m_lblk,
                        (unsigned)map->m_len, -err);

(3)Unmap the addr cause warning at

__set_page_dirty:WARN_ON_ONCE(warn && !PageUptodate(page));

(4) wait for a minute,then bugon happen.

Cc: stable@vger.kernel.org
Signed-off-by: wangguang <wangguang03@zte.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
fs/ext4/inode.c

index f058afbc4c46e10890bfed87728b5ed5d53805db..9b464e5272bbab69e5baf44eb031e07278e05b55 100644 (file)
@@ -1649,6 +1649,8 @@ static void mpage_release_unused_pages(struct mpage_da_data *mpd,
                        BUG_ON(!PageLocked(page));
                        BUG_ON(PageWriteback(page));
                        if (invalidate) {
+                               if (page_mapped(page))
+                                       clear_page_dirty_for_io(page);
                                block_invalidatepage(page, 0, PAGE_SIZE);
                                ClearPageUptodate(page);
                        }