Hi, On 07/08/2019 23:13, Pascal Van Leeuwen wrote: >> It is a bit confusing, but it is the only reasonable way to support >> variably sized context and at the same time keep the whole request in >> a single allocation. Yes, and the reason here it was detected only for aesni_intel is that it is submitted though more layers, these depends on variable context length (at least it run through crypt_simd layer). I think all other implementations on this 32bit machine were called directly, so no corruption seen there. > Ah, ok, I did not know anything about that ... so there's really no way > I could've done this correctly or to have found the problem myself really. > Good that it's resolved now, though. > > I fixed a couple of other minor things already, is it OK if I roll this > into an update to my original patch? Sure, feel free to fold in my change in v2 patch. I'll test it and provide Tested-by later. Maybe it would be good to also include /* must be the last */ comment to the req in struct, though. Also, maybe this req should have CRYPTO_MINALIGN_ATTR attribute also? I expect req can be run through exotic hw accelerators later with some memory alignment requirements. Ondra will know better here, though ;-) Thanks! Milan
