From mboxrd@z Thu Jan 1 00:00:00 1970 From: Roland Dreier Subject: Re: Re: [PATCH] infiniband/mthca : Fix userland mapping of mthca infiniband cards in Xen dom0 Date: Thu, 06 Jan 2011 10:51:57 -0800 Message-ID: References: <1293034260.30522.426.camel@trax> <20110106165549.GC18774@dumpdata.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: In-Reply-To: <20110106165549.GC18774@dumpdata.com> (Konrad Rzeszutek Wilk's message of "Thu, 6 Jan 2011 11:55:49 -0500") List-Unsubscribe: , List-Post: List-Help: List-Subscribe: , Sender: xen-devel-bounces@lists.xensource.com Errors-To: xen-devel-bounces@lists.xensource.com To: Konrad Rzeszutek Wilk Cc: jackm@mellanox.co.il, Jeremy Fitzhardinge , xen-devel@lists.xensource.com, rdreir@cisco.com, linux-rdma@vger.kernel.org, Vivien Bernet-Rollande , sean.hefty@intel.com List-Id: xen-devel@lists.xenproject.org > There is another way, which I hope can fix this issue (it fixes > the graphics drivers). Basically making the Xen MMU be aware of the E820 > and consider any access to PFNs that are not in RAM to have VM_IO implicitly set. > > The patches are at git://git.kernel.org/pub/scm/linux/kernel/git/konrad/xen.git > > stable/p2m-identity.v4.3 > > And posted on LKML )https://lkml.org/lkml/2010/12/30/163) > > They _should_ take care of your problem as long as the PFNs you > are mapping fall within the PCI BAR regions (which I would presume they > do since you are using ioremap). Yes, that should work fine for all the cases I know of in the RDMA drivers. We are just mapping some PCI BAR space into userspace for direct access to hardware (just like graphics, I think). I definitely prefer a solution that doesn't require driver authors to add a Xen-specific fix that isn't required on any other architecture. - R.