Hi: Dave, please don't apply this series of patches as it's going into cryptodev-2.6 :)
This series of patches adds a new type of operations (the existing ones are cipher, digest and compress). The idea is to differentiate between block chaining ciphers versus simple block ciphers from both a user point of view and an implementation point of view. >From the user's perspective, the block cipher type is really identical to the existing cipher type. However, the revamped cipher type is quite different in that it now offers just two main operations, encrypt_one and decrypt_one. Both of these take just a plain kernel pointer as opposed to a scatterlist. This greatly simplifies the places that actually need simple ciphers since a block is tiny (usually 8 to 16 bytes) and describing it using a scatterlist is really overkill. We never want to offload a simple cipher operation to hardware because of the tiny amount of data that's in a single block. Even if we did copying 8 bytes/16 bytes where necessary would be trivial in terms of cost. >From the algorithm writer's perspective, the block cipher type is quite different in that it presents and only presents an SG interface as opposed to the plain pointer (one or multiple blocks) interface provided by the current cipher type. This is one of the main motiviations since we need this for offload crypto. For software algorithms, the two existing block chaining algorithms (CBC/ECB) have been rewritten as a template that takes a simple cipher as a parameter. Incidentally, I took the opportunity to optimise the software CBC implementation which gives a 10% boost in case of in-place encryption and not-in-place decryption (by avoiding the copy of the IV). This puts them on a level playing field with other algorithms that implement block chaining operations, that is, padlock and s390. Because these all operate in the CPU address space, they need assistance to walk the scatterlists. These are provided in the form of blkcipher_walk_virt() and blkcipher_walk_done() . There is also blkcipher_walk_phys() which will be used by offload drivers since they do not require a virtual mapping of the input or output. The main user-visible change is that instead of allocating aes with a CBC mode argument you now allocate a block cipher object called cbc(aes). All existing users have been converted to this scheme. I've also taken the opportunity to start the move towards a more strongly typed crypto_tfm structure. Previously in time we relied on run-time checks to make sure that the crypto_tfm object supplied by the user is suitable for the operation requested. This is now done at compile-time for ciphers and block ciphers. Digest and compress will follow soon. I'm in the process of converting digest across to a new format (in fact it'll get a new name too -- "hash" due to the conversion process). That's why I've temporarily reverted the hmac template code that I posted before. It should return within a few days. Once the digest/compress conversion is complete, I will add the first asynchronous algorithm and then begin converting IPsec over. Cheers, -- Visit Openswan at http://www.openswan.org/ Email: Herbert Xu ~{PmV>HI~} <[EMAIL PROTECTED]> Home Page: http://gondor.apana.org.au/~herbert/ PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt - To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html
