On Fri, Oct 9, 2015 at 10:17 AM, Valentine Sinitsyn <valentine.sinit...@gmail.com> wrote: > Hi David, > > Thanks for your efforts. > > > On 09.10.2015 07:53, David Kiarie wrote: >> >> David (4): >> hw/core: Add iommu to machine properties >> hw/pci-host: Add AMD IOMMU to PIIX and Q35 pcs >> hw/i386: Introduce AMD IOMMU >> hw/acpi: ACPI table for AMD IOMMU >> >> hw/core/machine.c | 25 + >> hw/i386/Makefile.objs | 1 + >> hw/i386/acpi-build.c | 85 +++ >> hw/i386/amd_iommu.c | 1266 >> +++++++++++++++++++++++++++++++++++++++++++ >> hw/i386/amd_iommu.h | 363 +++++++++++++ >> hw/pci-host/piix.c | 10 + >> hw/pci-host/q35.c | 10 + >> include/hw/acpi/acpi-defs.h | 55 ++ >> include/hw/boards.h | 2 + >> 9 files changed, 1817 insertions(+) >> create mode 100644 hw/i386/amd_iommu.c >> create mode 100644 hw/i386/amd_iommu.h >> >> >> AMD IOMMU v1 >> >> Changes since RFC >> -code cleanup around MMIO code >> -compacted startup code - some of the 'init' and 'reset' code is >> similar and can be reused >> -Implemented basic IOTLB - it's just a hashtable of already requested >> virtual address translations whereby if a device requests translation of the >> same address we don't have to do a page walk, again. >> -Implemented event logging - this version logs about all possible >> SW/HW errors but most of them are very unlikely to occur. >> -Changed GVA and HVA to go by the physical host bus implementations - >> 48 bits virtual and 40 bits physical address space. > > I haven't looked into this closely yet, but shouldn't you support a whole 63 > bits wide addresses (and 6-level page tables), as per specification? > Am not sure too.
When I tested this, linux was using 3-level pages. I went with 48 bits as AFAIK that's what current hardware works. This code should work with 6-levels though I just used 48 to be 'safe'. >> >> TODO >> -Implement MMIO reserving code >> -Implement r/wc MMIO registers >> -Implement interrupts related to r/wc registers >> >> IOMMU won't be able to implement PPR logging with the current Qemu >> infrastructure - Qemu doesn't implement anything that's related to PASID. >> Similary, guest translation won't be possible. >> >> Note: I have not tested this code and it might have a few bugs (silly >> bugs) of course but the RFC version works so, if present, it's just minor >> bugs. > > > Regards, > Valentine Sinitsyn