Messages in this thread | | | Subject | Re: [PATCH] intel-iommu: Fix alloc_coherent for pass-through devices | From | Alex Williamson <> | Date | Thu, 22 Oct 2009 06:24:01 -0600 |
| |
On Thu, 2009-10-22 at 15:28 +0900, David Woodhouse wrote: > On Wed, 2009-10-21 at 21:41 -0600, Alex Williamson wrote: > > After 19943b0e (intel-iommu: Unify hardware and software passthrough > > support) hardware pass-through mode devices make use of intel_dma_ops > > rather than swiotlb_dma_ops. The problem is that intel_alloc_coherent > > ignores the device coherent_dma_mask when allocating the page since it > > expects to remap the page and provide the device with an iova within the > > coherent mask. This breaks when we use pass-through. > > > > The patch below crudely works around the problem, but I hope we can come > > up with something better without reintroducing the dependency on > > swiotlb. The device hitting this problem is an HP smart array > > controller on a Proliant G6 system. It uses a default 32bit coherent > > DMA mask, and stalls, presumably waiting on control data to change in > > the wrong address space, when it gets a coherent buffer above 4G. This > > device also doesn't exactly play nice when using VT-d in anything other > > than pass-through mode, so switching it into mapped mode is not really > > an option. > > I don't understand. If your device can't cope with 64-bit addresses, > then surely iommu_no_mapping() should return _false_ for it. And we'll > actually use the IOMMU even though we're generally in passthrough mode.
The coherent_dma_mask is independent of the dma_mask and can be set separately by the device. The default for any device that doesn't specify one is 32bits. iommu_should_identity_map() only checks the dma_mask, not the coherent_dma_mask. We never see "32bit %s uses non-identity mapping" for this device, and I suspect if we did it also wouldn't work because of the mentioned unfriendliness in mapped mode. Since coherent mappings are generally a slow path, I think we should be making more effort to get a usable buffer before dumping the device out of passthrough mode. BTW, we skip RMRR setup when doing hardware pass-through, but I can't find where they get reloaded if we then end up removing the device from the si_domain. Is this another issue? Thanks,
Alex
| |