From: Jeremy Fitzhardinge Subject: Re: Re: [PATCH 1 of 7] x86: add _PAGE_IOMAP pte flag for IO mappings Date: Tue, 09 Sep 2008 08:48:06 -0700 Message-ID: <48C69AB6.20500@goop.org> References: Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: List-Unsubscribe: , List-Post: List-Help: List-Subscribe: , Sender: xen-devel-bounces@lists.xensource.com Errors-To: xen-devel-bounces@lists.xensource.com To: Keir Fraser Cc: Xen Devel , linux-kernel@vger.kernel.org, Avi Kivity , Andi Kleen , "H. Peter Anvin" , Ingo Molnar List-Id: xen-devel@lists.xenproject.org Keir Fraser wrote: > That's basically what our _PAGE_IO flag (in our old Linux patchset) means. > We use it to cause pte_pfn() to return an invalid pfn and hence avoid > reference counting that way. Since kernel mappings are never reference > counted (I think?) perhaps we could use _PAGE_SPECIAL even if it is > restricted to use on user mappings. Well, _PAGE_IOMAP's most important semantic from Xen's perspective is that the frame number is considered to already be an MFN and so isn't converted. It may be that _PAGE_SPECIAL is also useful for its "no refcount" properties, but we could set both in that case. At present we can't set _PAGE_SPECIAL on kernel mappings because we assume its usermode only, and its shared with _PAGE_CPA_TEST which is kernel-only. Most _PAGE_IOMAP mappings are in kernel space, but I wouldn't discount user mappings if we end up allowing direct access to passthrough devices. J