nilfs2: fix data loss with mmap()

This bug leads to reproducible silent data loss, despite the use of
msync(), sync() and a clean unmount of the file system.  It is easily
reproducible with the following script:

  ----------------[BEGIN SCRIPT]--------------------
  mkfs.nilfs2 -f /dev/sdb
  mount /dev/sdb /mnt

  dd if=/dev/zero bs=1M count=30 of=/mnt/testfile

  umount /mnt
  mount /dev/sdb /mnt
  CHECKSUM_BEFORE="$(md5sum /mnt/testfile)"

  /root/mmaptest/mmaptest /mnt/testfile 30 10 5

  sync
  CHECKSUM_AFTER="$(md5sum /mnt/testfile)"
  umount /mnt
  mount /dev/sdb /mnt
  CHECKSUM_AFTER_REMOUNT="$(md5sum /mnt/testfile)"
  umount /mnt

  echo "BEFORE MMAP:\t$CHECKSUM_BEFORE"
  echo "AFTER MMAP:\t$CHECKSUM_AFTER"
  echo "AFTER REMOUNT:\t$CHECKSUM_AFTER_REMOUNT"
  ----------------[END SCRIPT]--------------------

The mmaptest tool looks something like this (very simplified, with
error checking removed):

  ----------------[BEGIN mmaptest]--------------------
  data = mmap(NULL, file_size - file_offset, PROT_READ | PROT_WRITE,
              MAP_SHARED, fd, file_offset);

  for (i = 0; i < write_count; ++i) {
        memcpy(data + i * 4096, buf, sizeof(buf));
        msync(data, file_size - file_offset, MS_SYNC))
  }
  ----------------[END mmaptest]--------------------

The output of the script looks something like this:

  BEFORE MMAP:    281ed1d5ae50e8419f9b978aab16de83  /mnt/testfile
  AFTER MMAP:     6604a1c31f10780331a6850371b3a313  /mnt/testfile
  AFTER REMOUNT:  281ed1d5ae50e8419f9b978aab16de83  /mnt/testfile

So it is clear, that the changes done using mmap() do not survive a
remount.  This can be reproduced a 100% of the time.  The problem was
introduced in commit 136e877 ("nilfs2: fix issue of
nilfs_set_page_dirty() for page at EOF boundary").

If the page was read with mpage_readpage() or mpage_readpages() for
example, then it has no buffers attached to it.  In that case
page_has_buffers(page) in nilfs_set_page_dirty() will be false.
Therefore nilfs_set_file_dirty() is never called and the pages are never
collected and never written to disk.

This patch fixes the problem by also calling nilfs_set_file_dirty() if the
page has no buffers attached to it.

[[email protected]: s/PAGE_SHIFT/PAGE_CACHE_SHIFT/]
Signed-off-by: Andreas Rohner <[email protected]>
Tested-by: Andreas Rohner <[email protected]>
Signed-off-by: Ryusuke Konishi <[email protected]>
Cc: <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>

fs/nilfs2/inode.c

@@ -24,6 +24,7 @@
 #include <linux/buffer_head.h>
 #include <linux/gfp.h>
 #include <linux/mpage.h>
+#include <linux/pagemap.h>
 #include <linux/writeback.h>
 #include <linux/aio.h>
 #include "nilfs.h"
@@ -219,10 +220,10 @@ static int nilfs_writepage(struct page *page, struct writeback_control *wbc)
 
 static int nilfs_set_page_dirty(struct page *page)
 {
+	struct inode *inode = page->mapping->host;
 	int ret = __set_page_dirty_nobuffers(page);
 
 	if (page_has_buffers(page)) {
-		struct inode *inode = page->mapping->host;
 		unsigned nr_dirty = 0;
 		struct buffer_head *bh, *head;
 
@@ -245,6 +246,10 @@ static int nilfs_set_page_dirty(struct page *page)
 
 		if (nr_dirty)
 			nilfs_set_file_dirty(inode, nr_dirty);
+	} else if (ret) {
+		unsigned nr_dirty = 1 << (PAGE_CACHE_SHIFT - inode->i_blkbits);
+
+		nilfs_set_file_dirty(inode, nr_dirty);
 	}
 	return ret;
 }