From: Tadeusz Struk Subject: Re: [PATCH] crypto: af_alg - add async support to algif_aead Date: Tue, 19 Jan 2016 07:18:32 -0800 Message-ID: <569E53C8.9090803@intel.com> References: <20160115192112.3065.45755.stgit@tstruk-mobl1> <1641011.7xuR0fALWB@myon.chronox.de> <569D034F.5090905@gmail.com> <20160119003428.GA5571@gondor.apana.org.au> Mime-Version: 1.0 Content-Type: text/plain; charset=windows-1252 Content-Transfer-Encoding: 7bit Cc: Stephan Mueller , linux-kernel@vger.kernel.org, linux-crypto@vger.kernel.org To: Herbert Xu , Tadeusz Struk Return-path: Received: from mga02.intel.com ([134.134.136.20]:26921 "EHLO mga02.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754297AbcASPWT (ORCPT ); Tue, 19 Jan 2016 10:22:19 -0500 In-Reply-To: <20160119003428.GA5571@gondor.apana.org.au> Sender: linux-crypto-owner@vger.kernel.org List-ID: On 01/18/2016 04:34 PM, Herbert Xu wrote: >> My understanding is that the sock_kmalloc is mainly used for allocations >> > of the user provided data, because it keeps tracks of how much memory >> > is allocated by a socket, and makes sure that is will not exceed the >> > sysctl_optmem_max limit. Usually the internal structures, with fixed >> > size are allocated simply with kmalloc. I don't think that using >> > sock_kmalloc will give us any benefit here. > If there is only ever one of them per-socket then kmalloc is fine, > otherwise you should use sock_kmalloc. There is one per request. There can be a few of them at a given time. We have the same thing in skcipher and we use kmalloc there. > >> I agree that they are very similar, but I found it much easier to debug >> > when they are separate functions. I would prefer to keep them separate. >> > They are also separate in algif_skcipher. It makes it also easier to >> > read and understand. > I too would prefer a common function. However we can do this > later if we wish. > lets do this later then. > >> > The inflight ctr is incremented only if an asynchronous request has been >> > successfully en-queued for processing. If a user forges to call recvmsg >> > then the function that increments it won't be even called. >> > >From the other hand we don't want to give the option to interrupt the >> > wait, because in a case, when we do have request being processed by some >> > hardware, and the user kills the process, causing the socket to be >> > freed, then we will get an Oops in the callback. > This should be replaced with a sock_hold. Ok, I will try sock_hold. Thanks, -- TS