From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1754229AbaIOQr5 (ORCPT ); Mon, 15 Sep 2014 12:47:57 -0400 Received: from foss-mx-na.foss.arm.com ([217.140.108.86]:33056 "EHLO foss-mx-na.foss.arm.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754073AbaIOQr4 (ORCPT ); Mon, 15 Sep 2014 12:47:56 -0400 Date: Mon, 15 Sep 2014 17:47:38 +0100 From: Catalin Marinas To: Jon Masters Cc: "linux-arm-kernel@lists.infradead.org" , "linux-kernel@vger.kernel.org" Subject: Re: "dma-coherent" property inheritance (arm vs arm64) Message-ID: <20140915164738.GJ5415@arm.com> References: <5414FF8F.5040409@redhat.com> <20140915155216.GG5415@arm.com> <54171485.5010803@redhat.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <54171485.5010803@redhat.com> User-Agent: Mutt/1.5.21 (2010-09-15) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Mon, Sep 15, 2014 at 05:32:05PM +0100, Jon Masters wrote: > On 09/15/2014 11:52 AM, Catalin Marinas wrote: > > Hi Jon, > > > > On Sun, Sep 14, 2014 at 03:38:07AM +0100, Jon Masters wrote: > >> commit 6ecba8eb51b7d23fda66388a5420be7d8688b186 > >> Author: Catalin Marinas > >> Date: Fri Apr 25 15:31:45 2014 +0100 > >> > >> arm64: Use bus notifiers to set per-device coherent DMA ops > >> > >> Thus at this point, on 32-bit systems, we have defined this function: > >> > >> set_arch_dma_coherent_ops > > > > It was a timing issue that they are not in sync. I would have used > > set_arch_dma_coherent_ops() but it wasn't very clear when it gets merged > > and changing the default on arm64 broke some other assumptions, so a > > temporary fix. > > Ok. So arm64 will automatically switch over to set_arch_dma_coherent_ops > (once it's there, it will be called automatically anyway because it's in > the callchain anyway already). I still need to figure out if set_arch_dma_coherent_ops() is called for AMBA devices. > For the moment, on our internal PCI trees, we'll add a notifier and wait > for upstream to get PCI support and figure this out as part of that > activity. In the case of ACPI, I'm still pushing for _CCA methods to be > added by default because it makes sense to be verbose in describing > platform hardware devices, even in spite of default. Anyway, with ACPI it would probably make sense to keep the hooks (with the recursive parent check), so set_arch_dma_coherent_ops() wouldn't buy us much (especially if it's only for platform devices). -- Catalin