[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [Xen-devel] [PATCH] PCI/MSI: don't disable AMD IOMMU MSI on Xen dom0
On 06/21/2012 02:06 PM, Jan Beulich wrote: Could you please provide a particular c/s number?... (I saw too many c/s might be related to this topic). so that I could work out a patch to support both i/o and mmcfg.On 21.06.12 at 13:21, Wei Wang<wei.wang2@xxxxxxx> wrote:On 06/21/2012 11:59 AM, Jan Beulich wrote:On 14.06.12 at 17:15, Wei Wang<wei.wang2@xxxxxxx> wrote:Am 14.06.2012 16:18, schrieb Jan Beulich:Have you at all considered getting this fixed on the kernel side? As I don't have direct access to any AMD IOMMU capable system - can one, other than by enumerating the respective PCI IDs or reading ACPI tables, reasonably easily identify the devices in question (e.g. via vendor/class/sub-class or some such)? That might permit skipping those in the offending kernel code...AMD IOMMUs (both v1 and v2) uses class id 08 (System Base Peripheral) and sub class id 06 (IOMMU). Combined with PCI_VENDEOR_ID_AMD, this should be enough to identify amd iommu device. I could send you a kernel patch for review using this approach. I would believe that fixing this issue in 4.2, no matter how, is really important for amd iommu.As you didn't come forward with anything, here's my first take on this:Hi Jan Thanks a lot for the patch. Actually I plan to send my version today, which is based on 3.4 pv_ops but looks very similar to yours. So, Acked! I also evaluated the possibility of hiding iommu device from dom0. I think the change is no quite a lot, at least, for io based pcicfg access. A proof-of-concept patch is attached.This completely hides the device from Dom0, but only when config space is accessed via method 1. Did you not see my earlier patch doing this for MCFG as well (albeit only disallowing Sounds better to me...this still allows user to check iommu status from lspci.writes, so allowing the device to still be seen by Dom0)? Whether completely hiding the device is actually okay I can't easily tell: Would IOMMUs always be either at func 0 of a single- unction device, or at a non-zero func of a multi-function one? If not, other devices may get hidden implicitly. AMD IOMMU is an independent pci-e endpoint, and this function will not be used for other purposes other than containing an iommu. So I don't see that iommu will share bdf value with other devices. Thanks, Wei Also I noticed just now that guest_io_read() wouldn't really behave correctly when pci_cfg_ok() returned false - it might pass back 0xff even for a multi-byte read. I'll send a fix shortly. Jan--- a/xen/arch/x86/traps.c Thu Jun 21 11:30:59 2012 +0200 +++ b/xen/arch/x86/traps.c Thu Jun 21 13:19:02 2012 +0200 @@ -73,6 +73,7 @@ #include<asm/hpet.h> #include<public/arch-x86/cpuid.h> #include<xsm/xsm.h> +#include<asm/hvm/svm/amd-iommu-proto.h> /* * opt_nmi: one of 'ignore', 'dom0', or 'fatal'. @@ -1686,10 +1687,19 @@ static int pci_cfg_ok(struct domain *d, { uint32_t machine_bdf; uint16_t start, end; + struct amd_iommu *iommu; + if (!IS_PRIV(d)) return 0; machine_bdf = (d->arch.pci_cf8>> 8)& 0xFFFF; + + for_each_amd_iommu ( iommu ) + { + if ( machine_bdf == iommu->bdf ) + return 0; + } + start = d->arch.pci_cf8& 0xFF; end = start + size - 1; if (xsm_pci_config_permission(d, machine_bdf, start, end, write)) _______________________________________________ Xen-devel mailing list Xen-devel@xxxxxxxxxxxxx http://lists.xen.org/xen-devel
|
Lists.xenproject.org is hosted with RackSpace, monitoring our |