fuse: postpone end_page_writeback() in fuse_writepage_locked()
Maxim Patlasov [Mon, 12 Aug 2013 16:39:30 +0000 (20:39 +0400)]
commit 4a4ac4eba1010ef9a804569058ab29e3450c0315 upstream.

The patch fixes a race between ftruncate(2), mmap-ed write and write(2):

1) An user makes a page dirty via mmap-ed write.
2) The user performs shrinking truncate(2) intended to purge the page.
3) Before fuse_do_setattr calls truncate_pagecache, the page goes to
   writeback. fuse_writepage_locked fills FUSE_WRITE request and releases
   the original page by end_page_writeback.
4) fuse_do_setattr() completes and successfully returns. Since now, i_mutex
   is free.
5) Ordinary write(2) extends i_size back to cover the page. Note that
   fuse_send_write_pages do wait for fuse writeback, but for another
   page->index.
6) fuse_writepage_locked proceeds by queueing FUSE_WRITE request.
   fuse_send_writepage is supposed to crop inarg->size of the request,
   but it doesn't because i_size has already been extended back.

Moving end_page_writeback to the end of fuse_writepage_locked fixes the
race because now the fact that truncate_pagecache is successfully returned
infers that fuse_writepage_locked has already called end_page_writeback.
And this, in turn, infers that fuse_flush_writepages has already called
fuse_send_writepage, and the latter used valid (shrunk) i_size. write(2)
could not extend it because of i_mutex held by ftruncate(2).

Signed-off-by: Maxim Patlasov <mpatlasov@parallels.com>
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

fs/fuse/file.c

index 35f2810..4943f2c 100644 (file)
@@ -1530,7 +1530,6 @@ static int fuse_writepage_locked(struct page *page)
 
        inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
        inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
-       end_page_writeback(page);
 
        spin_lock(&fc->lock);
        list_add(&req->writepages_entry, &fi->writepages);
@@ -1538,6 +1537,8 @@ static int fuse_writepage_locked(struct page *page)
        fuse_flush_writepages(inode);
        spin_unlock(&fc->lock);
 
+       end_page_writeback(page);
+
        return 0;
 
 err_free: