On Tue, Jan 03, 2017 at 11:33:27AM +0100, Alexander Gordeev wrote:
> On Mon, Dec 12, 2016 at 11:08:18AM +0800, Peter Xu wrote:
> > DMAR test is based on QEMU edu device. A 4B DMA memory copy is carried
> > out as the simplest DMAR test.
>
> Hi Peter,
>
> Sorry for missing to provide feedback on your series.
> I did just a cursory review, but did not notice few issue below.
>
> > Signed-off-by: Peter Xu <peterx@xxxxxxxxxx>
> > ---
> > lib/pci.h | 5 ++
> > lib/x86/intel-iommu.c | 133 ++++++++++++++++++++++++++++++++++++++++++++++++++
> > lib/x86/intel-iommu.h | 4 ++
> > x86/Makefile.common | 1 +
> > x86/intel-iommu.c | 51 +++++++++++++++++++
> > 5 files changed, 194 insertions(+)
> >
> > diff --git a/lib/pci.h b/lib/pci.h
> > index d3052ef..26968b1 100644
> > --- a/lib/pci.h
> > +++ b/lib/pci.h
> > @@ -18,6 +18,9 @@ enum {
> > #define PCI_BAR_NUM 6
> > #define PCI_DEVFN_MAX 256
> >
> > +#define PCI_BDF_GET_DEVFN(x) ((x) & 0xff)
> > +#define PCI_BDF_GET_BUS(x) (((x) >> 8) & 0xff)
> > +
> > struct pci_dev {
> > uint16_t bdf;
> > phys_addr_t resource[PCI_BAR_NUM];
> > @@ -28,6 +31,8 @@ extern void pci_scan_bars(struct pci_dev *dev);
> > extern void pci_cmd_set_clr(struct pci_dev *dev, uint16_t set, uint16_t clr);
> > extern void pci_enable_defaults(struct pci_dev *dev);
> >
> > +typedef phys_addr_t iova_t;
> > +
> > extern bool pci_probe(void);
> > extern void pci_print(void);
> > extern bool pci_dev_exists(pcidevaddr_t dev);
> > diff --git a/lib/x86/intel-iommu.c b/lib/x86/intel-iommu.c
> > index 1887998..86b1805 100644
> > --- a/lib/x86/intel-iommu.c
> > +++ b/lib/x86/intel-iommu.c
> > @@ -11,6 +11,42 @@
> > */
> >
> > #include "intel-iommu.h"
> > +#include "libcflat.h"
> > +
> > +/*
> > + * VT-d in QEMU currently only support 39 bits address width, which is
> > + * 3-level translation.
> > + */
> > +#define VTD_PAGE_LEVEL 3
> > +#define VTD_CE_AW_39BIT 0x1
> > +
> > +typedef uint64_t vtd_pte_t;
> > +
> > +struct vtd_root_entry {
> > + /* Quad 1 */
> > + uint64_t present:1;
> > + uint64_t __reserved:11;
> > + uint64_t context_table_p:52;
> > + /* Quad 2 */
> > + uint64_t __reserved_2;
> > +} __attribute__ ((packed));
> > +typedef struct vtd_root_entry vtd_re_t;
> > +
> > +struct vtd_context_entry {
> > + /* Quad 1 */
> > + uint64_t present:1;
> > + uint64_t disable_fault_report:1;
> > + uint64_t trans_type:2;
> > + uint64_t __reserved:8;
> > + uint64_t slptptr:52;
> > + /* Quad 2 */
> > + uint64_t addr_width:3;
> > + uint64_t __ignore:4;
> > + uint64_t __reserved_2:1;
> > + uint64_t domain_id:16;
> > + uint64_t __reserved_3:40;
> > +} __attribute__ ((packed));
> > +typedef struct vtd_context_entry vtd_ce_t;
> >
> > #define VTD_RTA_MASK (PAGE_MASK)
> > #define VTD_IRTA_MASK (PAGE_MASK)
> > @@ -83,6 +119,103 @@ static void vtd_setup_ir_table(void)
> > printf("IR table address: 0x%016lx\n", vtd_ir_table());
> > }
> >
> > +static void vtd_install_pte(vtd_pte_t *root, iova_t iova,
> > + phys_addr_t pa, int level_target)
> > +{
> > + int level;
> > + unsigned int offset;
> > + void *page;
> > +
> > + for (level = VTD_PAGE_LEVEL; level > level_target; level--) {
> > + offset = PGDIR_OFFSET(iova, level);
> > + if (!(root[offset] & VTD_PTE_RW)) {
> > + page = alloc_page();
> > + memset(page, 0, PAGE_SIZE);
> > + root[offset] = virt_to_phys(page) | VTD_PTE_RW;
> > + }
> > + root = (uint64_t *)(phys_to_virt(root[offset] &
> > + VTD_PTE_ADDR));
> > + }
> > +
> > + offset = PGDIR_OFFSET(iova, level);
> > + root[offset] = pa | VTD_PTE_RW;
> > + if (level != 1) {
> > + /* This is huge page */
> > + root[offset] |= VTD_PTE_HUGE;
> > + }
> > +}
> > +
> > +#define VTD_PHYS_TO_VIRT(x) \
> > + ((void *)(((uint64_t)phys_to_virt(x)) >> VTD_PAGE_SHIFT))
>
> I would say this macro is rather superfluous.
>
> > +/**
> > + * vtd_map_range: setup IO address mapping for specific memory range
> > + *
> > + * @sid: source ID of the device to setup
> > + * @iova: start IO virtual address
> > + * @pa: start physical address
> > + * @size: size of the mapping area
> > + */
> > +void vtd_map_range(uint16_t sid, iova_t iova, phys_addr_t pa, size_t size)
> > +{
> > + uint8_t bus_n, devfn;
> > + void *slptptr;
> > + vtd_ce_t *ce;
> > + vtd_re_t *re = phys_to_virt(vtd_root_table());
> > +
> > + assert(IS_ALIGNED(iova, SZ_4K));
> > + assert(IS_ALIGNED(pa, SZ_4K));
> > + assert(IS_ALIGNED(size, SZ_4K));
> > +
> > + bus_n = PCI_BDF_GET_BUS(sid);
> > + devfn = PCI_BDF_GET_DEVFN(sid);
> > +
> > + /* Point to the correct root entry */
> > + re += bus_n;
> > +
> > + if (!re->present) {
> > + ce = alloc_page();
> > + memset(ce, 0, PAGE_SIZE);
> > + memset(re, 0, sizeof(*re));
> > + re->context_table_p = virt_to_phys(ce) >> VTD_PAGE_SHIFT;
> > + re->present = 1;
> > + printf("allocated vt-d root entry for PCI bus %d\n",
> > + bus_n);
> > + } else
> > + ce = VTD_PHYS_TO_VIRT(re->context_table_p);
>
> Seems, it should be instead something like:
>
> ce = phys_to_virt(re->context_table_p << VTD_PAGE_SHIFT);
Oops, you are right. It didn't encounter issue since current DMAR test
is too simple and only allocated one single page. :-)
I'll post a fix soon. Thanks!
-- peterx
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
