On Wed, Nov 23, 2011 at 03:43:05PM +0000, David Woodhouse wrote: > On Wed, 2011-11-23 at 16:41 +0100, Daniel Vetter wrote: > > Hm, that comment confuses me a bit. I've always thought that igfx_off only > > instantiates a identity mapping and leaves the dmar unit on. Is that > > wrong? > > It's completely off. If a DMAR unit has *only* graphics devices behind > it (as the one for the built-in graphics does, because it's a whole > boatload of speshul case for integration with the GTT), we just don't > enable it at all. See the second for_each_drhd_unit() loop in > init_no_remapping_devices(). That explanation confused me even more. So I've rechecked with intel_iommu=igfx_off and already thought I've made a decent fool of myself because I couldn't readily hang it. Turns out it's just much harder to hang with that. So I think moving around the tlb flushing for other dmar nodes to align with the idled igfx isn't a great solution, simply because I can't reliably tell whether it fixes anything. -Daniel -- Daniel Vetter Mail: daniel at ffwll.ch Mobile: +41 (0)79 365 57 48
