[BUG] Rewriting backing_dev_info in MTD

[Kirill A. Shutemov]

I've got NULL-pointer dereference in __mark_inode_dirty() on chmod()
for MTD device node. wb->bdi was NULL in this case.

During investigation I've found that MTD subsystem rewrites
file->f_mapping->backing_dev_info on openning to get mmap() work on
MMU-less systems. But in fact it rewrites
inode->i_mapping->backing_dev_info too, since inode->i_mapping ==
file->f_mapping (see __dentry_open() in fs/open.c). It breaks
writeback of inode changes.

I guess the right way to fix this is changing of __dentry_open() to
create _copy_ of i_mapping to assign to f_mapping since in common case
f_mapping != i_mapping. But I'm not sure were the copy should be
freed.

What do you think?