From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([2001:4830:134:3::10]:51511) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1WDL4b-0002c3-Om for qemu-devel@nongnu.org; Tue, 11 Feb 2014 16:41:06 -0500 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1WDL4T-0003n4-Bs for qemu-devel@nongnu.org; Tue, 11 Feb 2014 16:40:57 -0500 Received: from mail-ea0-x22a.google.com ([2a00:1450:4013:c01::22a]:38412) by eggs.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1WDL4T-0003mn-4e for qemu-devel@nongnu.org; Tue, 11 Feb 2014 16:40:49 -0500 Received: by mail-ea0-f170.google.com with SMTP id g15so2197779eak.15 for ; Tue, 11 Feb 2014 13:40:48 -0800 (PST) Sender: Paolo Bonzini Message-ID: <52FA98DA.8030808@redhat.com> Date: Tue, 11 Feb 2014 22:40:42 +0100 From: Paolo Bonzini MIME-Version: 1.0 References: <1391420690-23745-1-git-send-email-edgar.iglesias@gmail.com> <52F7834B.8060004@suse.de> <20140210231055.GA25819@xsjedgari30.xlnx.xilinx.com> <20140210235315.GB25819@xsjedgari30.xlnx.xilinx.com> In-Reply-To: <20140210235315.GB25819@xsjedgari30.xlnx.xilinx.com> Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Subject: Re: [Qemu-devel] [PATCH v4 00/22] Steps towards per CPU address-spaces List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: "Edgar E. Iglesias" , Peter Maydell Cc: QEMU Developers , Blue Swirl , Anthony Liguori , pcrost@xilinx.com, =?ISO-8859-1?Q?Andreas_F=E4rber?= , Aurelien Jarno , Richard Henderson Il 11/02/2014 00:53, Edgar E. Iglesias ha scritto: >> > Is there that much overhead to creating an AS per master? I guess >> > you end up recalculating the same memory flatview for every AS >> > when a memory region gets mapped or unmapped. >> > In any case, if it's OK to share AddressSpaces between masters >> > then we can relatively easily do so, since there's already a list >> > of all AddressSpaces that you can run through to find the one >> > whose root MemoryRegion is the one you want. (Probably needs >> > a refcount in the AddressSpace struct as well.) > Right, something like that might be worthwhile. > > It might not make much difference until we start ramping > things up significantly (in terms of amount of masters). We already have a decent amount of masters for PCI, where each device that can do bus-master DMA has basically its own copy of address_space_memory. It has never been a problem in terms of time spent updating the ASes. Paolo