On Mon, Nov 04, 2019 at 06:01:17PM -0800, Eric Biggers wrote: > I think that "Severely bloating the per-I/O data structure" is an exaggeration, > since that it's only 32 bytes, and it isn't in struct bio directly but rather in > struct bio_crypt_ctx... Yes, and none of that is needed for the real inline crypto. And I think we can further reduce the overhead of bio_crypt_ctx once we have the basiscs sorted out. If we want to gain more traction we need to reduce the I/O to a minimum. > In any case, Satya, it might be a good idea to reorganize this patchset so that > it first adds all logic that's needed for "real" inline encryption support > (including the needed parts of blk-crypto.c), then adds the crypto API fallback > as a separate patch. That would separate the concerns more cleanly and make the > patchset easier to review, and make it easier to make the fallback > de-configurable or even remove it entirely if that turns out to be needed. Yes, that is a good idea. Not just in terms of patch, but also in terms of code organization. The current structure is pretty weird with 3 files that are mostly tighly integrated, except that one also has the software implementations. So what I think we need at a minimum is: - reoranizize that we have say block/blk-crypt.c for all the inline crypto infrastructure, and block/blk-crypy-sw.c for the actual software crypto implementation. - remove all the fields only needed for software crypto from bio_crypt_ctx, and instead clone the bio into a bioset with the additional fields only when we use the software implementation, so that there is no overhead for the hardware path.
