> -----Original Message----- > From: linux-crypto-owner@xxxxxxxxxxxxxxx <linux-crypto-owner@xxxxxxxxxxxxxxx> On Behalf Of Eric Biggers > Sent: Tuesday, January 28, 2020 10:13 PM > To: Gilad Ben-Yossef <gilad@xxxxxxxxxxxxx> > Cc: Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx>; Stephan Mueller <smueller@xxxxxxxxxx>; Linux Crypto Mailing List <linux- > crypto@xxxxxxxxxxxxxxx>; Geert Uytterhoeven <geert@xxxxxxxxxxxxxx>; David Miller <davem@xxxxxxxxxxxxx>; Ofir Drang > <Ofir.Drang@xxxxxxx> > Subject: Re: Possible issue with new inauthentic AEAD in extended crypto tests > > <<< External Email >>> > CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you recognize the > sender/sender address and know the content is safe. > > > On Tue, Jan 28, 2020 at 09:24:25AM +0200, Gilad Ben-Yossef wrote: > > - The source is presumed to have enough room for both the associated > > data and the plaintext. > > - Unless it's in-place encryption, in which case, you also presume to > > have room for the authentication tag > > The authentication tag is part of the ciphertext, not the plaintext. So the > rule is just that the ciphertext buffer needs to have room for it, not the > plaintext. > > Of course, when doing in-place encryption/decryption, the two buffers are the > same, so both will have room for it, even though the tag is only meaningful on > the ciphertext side. That's just the logical consequence of "in-place". > > > - The only way to tell if this is in-place encryption or not is to > > compare the pointers to the source and destination - there is no flag. > > Requiring users to remember to provide a flag to indicate in-place > encryption/decryption, in addition to passing the same scatterlist, would make > the API more complex. > Also, what would the benefit? You'd still have to compare the flag. The performance difference of comparing the flag vs comparing 2 pointers (that you need to read anyway) is likely completely negligible on most modern CPU architectures ... > > - You can count on the scattergather list not having a first NULL > > buffer, *unless* the plaintext and associated data length are both > > zero AND it's not in place encryption. > > - You can count on not getting NULL as a scatterlist point, *unless* > > the plaintext and associated data length are both zero AND it's not in > > place encryption. (I'm actually unsure of this one?) > > If we consider that the input is not just a scatterlist, but rather a > scatterlist and a length, then these observations are really just "you can > access the first byte, unless the length is 0" -- which is sort of obvious. And > requiring a dereferencable pointer for length = 0 is generally considered to be > bad API design; see the memcpy() fiasco > (https://www.imperialviolet.org/2016/06/26/nonnull.html). > > The API could be simplified by only supporting full scatterlists, but it seems > that users are currently relying on being able to encrypt/decrypt just a prefix. > > IMO, the biggest problems with the AEAD API are actually things you didn't > mention, such as the fact that the AAD isn't given in a separate scatterlist, > While I can understand this may be beneficial in some cases, I believe they do not outweigh the downsides: - In many use cases, AAD+cipher text are stored as one contiguous string. Requiring this string to be spit into seperate particles for AAD and ciphertext would be a burden. - For hardware accelerators, there is a cost associated with each additional particle, in terms of either bandwidth or performance or both. So less particles = better, generally. The only thing that I find odd is that if you do a non-inplace operation you have this undefined(?) gap in the output data where the AAD would be for inplace. That makes little sense to me and requires extra effort to skip over in the driver. > and that the API only supports scatterlists and not virtual addresses (which > makes it difficult to use in some cases). > While I can understand that this is difficult if the API user just got this virtual address provided from somewhere else and needs to do the translation, the other side of the medal is that any hardware driver would otherwise have to do address translation and scatterlist building on the fly (as hardware needs to access contiguous physical memory), which would be real burden there. While many API users_are_ able to provide a nice scatterlist at negligible extra cost. So why burden those? > In any case we do need much better documentation. I'm planning to improve some > of the crypto API documentation, but I'll probably do the hash and skcipher > algorithm types first before getting to AEAD. So if you want to improve the > AEAD documentation in the mean time, please go ahead. > > - Eric Regards, Pascal van Leeuwen Silicon IP Architect Multi-Protocol Engines, Rambus Security Rambus ROTW Holding BV +31-73 6581953 Note: The Inside Secure/Verimatrix Silicon IP team was recently acquired by Rambus. Please be so kind to update your e-mail address book with my new e-mail address. ** This message and any attachments are for the sole use of the intended recipient(s). It may contain information that is confidential and privileged. If you are not the intended recipient of this message, you are prohibited from printing, copying, forwarding or saving it. Please delete the message and attachments and notify the sender immediately. ** Rambus Inc.<http://www.rambus.com>
