Re: [RFC v4 2/4] crypto: Introduce CRYPTO_ALG_BULK flag

[Baolin Wang]

On 15 June 2016 at 14:49, Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx> wrote:
> On Wed, Jun 15, 2016 at 02:27:04PM +0800, Baolin Wang wrote:
>>
>> After some investigation, I still think we should divide the bulk
>> request from dm-crypt into small request (each one is 512bytes) if
>> this algorithm is not support bulk mode (like CBC). We have talked
>> with dm-crypt
>> maintainers why dm-crypt always use 512 bytes as one request size in
>> below thread, could you please check it?
>> http://www.kernelhub.org/?p=2&msg=907022
>
> That link only points to an email about an oops.

Ah, sorry. Would you check this thread?
http://lkml.iu.edu/hypermail/linux/kernel/1601.1/03829.html

>
> Diggin through that thread, the only objection I have seen is about
> the fact that you have to generate a fresh IV for each sector, which
> is precisely what I'm suggesting that you do.
>
> IOW, implement the IV generators in the crypto API, and then you can
> easily generate a new IV (if necessary) for each sector.
>
>> That means if we move the IV handling into crypto API, we still can
>> not use bulk interface for all algorithm, for example we still need to
>> read/write with 512 bytes for CBC, you can't use 4k or more block on
>> CBC (and most other encryption modes). If only a part of 4k block is
>> written (and then system crash happens), CBC would corrupt the block
>> completely. It means if we map one whole bio with bulk interface in
>> dm-crypt, we need to divide into every 512 bytes requests in crypto
>> layer. So I don't think we can handle every algorithm with bulk
>> interface just moving the IV handling into crypto API. Thanks.
>
> Of course you would do CBC in 512-byte blocks, but my point is that
> you should do this in a crypto API algorithm, rather than dm-crypt
> as we do now.  Once you implement that then dm-crypt can treat
> every algorithm as if they supported bulk processing.

But that means we should divide the bulk request into 512-byte size
requests and break up the mapped sg table for each request. Another
hand we should allocate memory for each request in crypto layer, which
dm-crypt have supplied one high efficiency way. I think these are
really top level how to use the crypro APIs, does that need to move
into crypto laryer? Thanks.

>
> Cheers,
> --
> Email: Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx>
> Home Page: http://gondor.apana.org.au/~herbert/
> PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt



-- 
Baolin.wang
Best Regards