Re: [PATCH 1/5] AF_RXRPC: Add blkcipher accessors for using kernel data directly [try #2]

[David Howells <dhowells@redhat.com>]

Christoph Hellwig <hch@infradead.org> wrote:

> I don't quite understand all these indirections.  What's the problem
> with just having a helper that builds the scatterlist for you?

I was trying to avoid building a scatterlist completely.  There's not much
point because the scatterlist approach involves finding out the page struct and
then kmapping it just so that the FCrypt algorithm can read 8 or 16 bytes of
data from kernel space.  Why do that if we can avoid it?  It's a waste of
processing time, and has to be done on every secure packet.

> We allow dma access to arbitary pieces of _dynamically_ allocated kernel
> memory, and I think using the crypto subsystem on the stack is not allowed
> at all.

FCrypt is only available in software as far as I know.  For producing and
checking packet signatures, using hardware support would be a waste of time as
the size of the crunched data is so small (a single 8-byte fragment per
packet).

> But the even bigger question is, how does this relate to rxrpc?

RxRPC has security features, up to and including full packet content encryption
if you select it.

> very odd line split

It's not really odd.  The "static" and "inline" don't actually add anything to
the function template.  They're merely scope limiters / optimisation
controllers, and so make a lot of sense placed on their own line.

David