Re: [PATCH 01/26] fscrypto: Add buffer operations

[Theodore Ts'o]

On Fri, Oct 21, 2016 at 06:24:00AM -0700, Christoph Hellwig wrote:
> > Another reason why I did the buffer functions is because fscrypt_encrypt_page()
> > always allocates a bounce page as temporary memory. For ext4 this is needed,
> > for UBIFS not.
> > UBIFS has already a construction buffer, especially since it also does compression.
> 
> We should defintively find a way to avoid that, but it's a separate
> issue from adding another API just to pass buffers.

Hmm, one approach we could use is to avoid allocating a bounce page if
the passed-in plaintext_page has the PageSlab flag set.  That would
work for ubifs, but if there are file systems that are using
get_free_page() for their particular construction buffer. it wouldn't
work for them.

Perhaps more importantly, are you planning on making compression +
encryption work?  Some security purists will say that compression +
encryption will leak some information about the plaintext (which is
technically true, but it's much like the people who don't want make it
easy to discard + encrypt, which Linus recently railed against).  So
my take is that as long as users understand that there are minor
leakage issues with compression + encryption, we should let them do
that --- and that would be an argument for supporting buffer
operations, and only requiring that the buffer size must be a multiple
of the underlying encryption block size.

						- Ted