On Tue, Jul 11, 2017 at 11:53:11AM -0700, Dave Watson wrote: > On 07/11/17 08:29 AM, Steffen Klassert wrote: > > Sorry for replying to old mail... > > > +int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx) > > > +{ > > > > ... > > > > > + > > > + if (!sw_ctx->aead_send) { > > > + sw_ctx->aead_send = crypto_alloc_aead("gcm(aes)", 0, 0); > > > + if (IS_ERR(sw_ctx->aead_send)) { > > > + rc = PTR_ERR(sw_ctx->aead_send); > > > + sw_ctx->aead_send = NULL; > > > + goto free_rec_seq; > > > + } > > > + } > > > + > > > > When I look on how you allocate the aead transformation, it seems > > that you should either register an asynchronous callback with > > aead_request_set_callback(), or request for a synchronous algorithm. > > > > Otherwise you will crash on an asynchronous crypto return, no? > > The intention is for it to be synchronous, and gather directly from > userspace buffers. It looks like calling > crypto_alloc_aead("gcm(aes)", 0, CRYPTO_ALG_ASYNC) is the correct way > to request synchronous algorithms only? Yes, but then you loose the aes-ni based algorithms because they are asynchronous. If you want to have good crypto performance, it is better to implement the asynchronous callbacks. > > > Also, it seems that you have your scatterlists on a per crypto > > transformation base istead of per crypto request. Is this intentional? > > We hold the socket lock and only one crypto op can happen at a time, > so we reuse the scatterlists. This is OK as long as the crypto happens synchronous. But as said above, I think this is not what you want.