Re: [f2fs-dev] [PATCH V2 10/13] fscrypt_encrypt_page: Loop across all blocks mapped by a page range

[Eric Biggers <ebiggers@kernel.org>]

Hi Chandan,

On Thu, May 02, 2019 at 11:22:05AM +0530, Chandan Rajendra wrote:
> On Thursday, May 2, 2019 3:59:01 AM IST Eric Biggers wrote:
> > Hi Chandan,
> > 
> > On Wed, May 01, 2019 at 08:19:35PM +0530, Chandan Rajendra wrote:
> > > On Wednesday, May 1, 2019 4:38:41 AM IST Eric Biggers wrote:
> > > > On Tue, Apr 30, 2019 at 10:11:35AM -0700, Eric Biggers wrote:
> > > > > On Sun, Apr 28, 2019 at 10:01:18AM +0530, Chandan Rajendra wrote:
> > > > > > For subpage-sized blocks, this commit now encrypts all blocks mapped by
> > > > > > a page range.
> > > > > > 
> > > > > > Signed-off-by: Chandan Rajendra <chandan@linux.ibm.com>
> > > > > > ---
> > > > > >  fs/crypto/crypto.c | 37 +++++++++++++++++++++++++------------
> > > > > >  1 file changed, 25 insertions(+), 12 deletions(-)
> > > > > > 
> > > > > > diff --git a/fs/crypto/crypto.c b/fs/crypto/crypto.c
> > > > > > index 4f0d832cae71..2d65b431563f 100644
> > > > > > --- a/fs/crypto/crypto.c
> > > > > > +++ b/fs/crypto/crypto.c
> > > > > > @@ -242,18 +242,26 @@ struct page *fscrypt_encrypt_page(const struct inode *inode,
> > > > > 
> > > > > Need to update the function comment to clearly explain what this function
> > > > > actually does now.
> > > > > 
> > > > > >  {
> > > > > >  	struct fscrypt_ctx *ctx;
> > > > > >  	struct page *ciphertext_page = page;
> > > > > > +	int i, page_nr_blks;
> > > > > >  	int err;
> > > > > >  
> > > > > >  	BUG_ON(len % FS_CRYPTO_BLOCK_SIZE != 0);
> > > > > >  
> > > > > 
> > > > > Make a 'blocksize' variable so you don't have to keep calling i_blocksize().
> > > > > 
> > > > > Also, you need to check whether 'len' and 'offs' are filesystem-block-aligned,
> > > > > since the code now assumes it.
> > > > > 
> > > > > 	const unsigned int blocksize = i_blocksize(inode);
> > > > > 
> > > > >         if (!IS_ALIGNED(len | offs, blocksize))
> > > > >                 return -EINVAL;
> > > > > 
> > > > > However, did you check whether that's always true for ubifs?  It looks like it
> > > > > may expect to encrypt a prefix of a block, that is only padded to the next
> > > > > 16-byte boundary.
> > > > > 		
> > > > > > +	page_nr_blks = len >> inode->i_blkbits;
> > > > > > +
> > > > > >  	if (inode->i_sb->s_cop->flags & FS_CFLG_OWN_PAGES) {
> > > > > >  		/* with inplace-encryption we just encrypt the page */
> > > > > > -		err = fscrypt_do_page_crypto(inode, FS_ENCRYPT, lblk_num, page,
> > > > > > -					     ciphertext_page, len, offs,
> > > > > > -					     gfp_flags);
> > > > > > -		if (err)
> > > > > > -			return ERR_PTR(err);
> > > > > > -
> > > > > > +		for (i = 0; i < page_nr_blks; i++) {
> > > > > > +			err = fscrypt_do_page_crypto(inode, FS_ENCRYPT,
> > > > > > +						lblk_num, page,
> > > > > > +						ciphertext_page,
> > > > > > +						i_blocksize(inode), offs,
> > > > > > +						gfp_flags);
> > > > > > +			if (err)
> > > > > > +				return ERR_PTR(err);
> > > > 
> > > > Apparently ubifs does encrypt data shorter than the filesystem block size, so
> > > > this part is wrong.
> > > > 
> > > > I suggest we split this into two functions, fscrypt_encrypt_block_inplace() and
> > > > fscrypt_encrypt_blocks(), so that it's conceptually simpler what each function
> > > > does.  Currently this works completely differently depending on whether the
> > > > filesystem set FS_CFLG_OWN_PAGES in its fscrypt_operations, which is weird.
> > > > 
> > > > I also noticed that using fscrypt_ctx for writes seems to be unnecessary.
> > > > AFAICS, page_private(bounce_page) could point directly to the pagecache page.
> > > > That would simplify things a lot, especially since then fscrypt_ctx could be
> > > > removed entirely after you convert reads to use read_callbacks_ctx.
> > > > 
> > > > IMO, these would be worthwhile cleanups for fscrypt by themselves, without
> > > > waiting for the read_callbacks stuff to be finalized.  Finalizing the
> > > > read_callbacks stuff will probably require reaching a consensus about how they
> > > > should work with future filesystem features like fsverity and compression.
> > > > 
> > > > So to move things forward, I'm considering sending out a series with the above
> > > > cleanups for fscrypt, plus the equivalent of your patches:
> > > > 
> > > > 	"fscrypt_encrypt_page: Loop across all blocks mapped by a page range"
> > > > 	"fscrypt_zeroout_range: Encrypt all zeroed out blocks of a page"
> > > > 	"Add decryption support for sub-pagesized blocks" (fs/crypto/ part only)
> > > > 
> > > > Then hopefully we can get all that applied for 5.3 so that fs/crypto/ itself is
> > > > ready for blocksize != PAGE_SIZE; and get your changes to ext4_bio_write_page(),
> > > > __ext4_block_zero_page_range(), and ext4_block_write_begin() applied too, so
> > > > that ext4 is partially ready for encryption with blocksize != PAGE_SIZE.
> > > > 
> > > > Then only the read_callbacks stuff will remain, to get encryption support into
> > > > fs/mpage.c and fs/buffer.c.  Do you think that's a good plan?
> > > 
> > > Hi Eric,
> > > 
> > > IMHO, I will continue posting the next version of the current patchset and if
> > > there are no serious reservations from FS maintainers the "read callbacks"
> > > patchset can be merged. In such a scenario, the cleanups being
> > > non-complicated, can be merged later.
> > > 
> > 
> > Most of the patches I have in mind are actually things that are in your patchset
> > already, or have been requested, or will be requested eventually :-).  I'm
> > concerned that people will keep going back and forth on this patchset for a lot
> > longer, arguing about fsverity, compression, details of the fs/crypto/ stuff,
> > etc.  Moreover it's based on unmerged patches that add the fsverity feature, so
> > it can't be merged as-is anyway.
> > 
> > IMO, it's also difficult for people to review the read_callbacks stuff when it's
> > mixed in with lots of other fscrypt and ext4 changes for blocksize != PAGE_SIZE.
> > 
> > I actually have a patchset almost ready already, so I'm going to send it out and
> > see what you think.  It *should* make things a lot easier for you, since then
> > you can base a much smaller read_callbacks patchset on top of it.
> 
> One of the things that I am concerned most about is the fact that the more we
> delay merging read_callbacks patchset, the more the chances of filesystems
> adding further operations that get executed after read I/O completes. Most of
> the time, these implementations tend to have filesystem specific changes which
> are going to be very difficult (impossible?) to make them work with
> read_callback patchset. So instead of making things easier, delaying merging
> the read_callback patchset ends up actually having the opposite effect.
> 
> With the read_callback patchset merged, FS feature developers will take
> read_callback framework into consideration before designing/implementing new
> related features.
> 

The main problems are that your patchset mixes up conceptually unrelated
changes, and is dependent on future filesystem features.  See how it starts by
adding read_callbacks support for both fscrypt *and* fsverity (the latter of
which is not merged yet), then updates fs/crypto/ to support subpage blocks,
*then* goes back and finishes read_callbacks to support buffer_heads since that
depended on the fs/crypto/ changes.  The ext4 changes for subpage blocks are
mixed in too throughout the patchset.  So I don't think it can proceed in its
current form; it's too much for anyone to handle at once.

And I see your first patchset for ext4 encryption with subpage blocks was sent
almost a year and a half ago, so it's indeed been going in circles for a while.

But based on your work I've been able to get the fs/crypto/ and ext4
preparations for subpage blocks into a clean set of changes by themselves.
There are needed in any case, so IMO we should take them first in order to
unblock the rest.

I don't really understand your point about forcing filesystems to be compatible
with read_callbacks.  The whole point of read_callbacks is that it's a common
support layer which makes it easier for filesystems to do the things they're
doing anyway, or will be doing.  So it shouldn't affect filesystem designs.

Thanks!

- Eric