Re: [Qemu-devel] [v2][PATCH 8/8] xen, gfx passthrough: add opregion mapping

[Stefano Stabellini]

On Fri, 16 May 2014, Tiejun Chen wrote:
> The OpRegion shouldn't be mapped 1:1 because the address in the host
> can't be used in the guest directly.
> 
> This patch traps read and write access to the opregion of the Intel
> GPU config space (offset 0xfc).
> 
> The original patch is from Jean Guyader <address@hidden>
> 
> Signed-off-by: Yang Zhang <address@hidden>
> Signed-off-by: Tiejun Chen <address@hidden>
> Cc: Jean Guyader <address@hidden>
> ---
> v2:
> 
> * We should return zero as an invalid address value while calling
>   igd_read_opregion().
> 
>  hw/xen/xen_pt.h             |  4 +++-
>  hw/xen/xen_pt_config_init.c | 45 ++++++++++++++++++++++++++++++++++++++++++-
>  hw/xen/xen_pt_graphics.c    | 47 
> +++++++++++++++++++++++++++++++++++++++++++++
>  3 files changed, 94 insertions(+), 2 deletions(-)
> 
> diff --git a/hw/xen/xen_pt.h b/hw/xen/xen_pt.h
> index 507165c..25147cf 100644
> --- a/hw/xen/xen_pt.h
> +++ b/hw/xen/xen_pt.h
> @@ -63,7 +63,7 @@ typedef int (*xen_pt_conf_byte_read)
>  #define XEN_PT_BAR_UNMAPPED (-1)
>  
>  #define PCI_CAP_MAX 48
> -
> +#define PCI_INTEL_OPREGION 0xfc
>  
>  typedef enum {
>      XEN_PT_GRP_TYPE_HARDWIRED = 0,  /* 0 Hardwired reg group */
> @@ -306,5 +306,7 @@ int pci_create_pch(PCIBus *bus);
>  void igd_pci_write(PCIDevice *pci_dev, uint32_t config_addr,
>                     uint32_t val, int len);
>  uint32_t igd_pci_read(PCIDevice *pci_dev, uint32_t config_addr, int len);
> +uint32_t igd_read_opregion(XenPCIPassthroughState *s);
> +void igd_write_opregion(XenPCIPassthroughState *s, uint32_t val);
>  
>  #endif /* !XEN_PT_H */
> diff --git a/hw/xen/xen_pt_config_init.c b/hw/xen/xen_pt_config_init.c
> index de9a20f..cf36a40 100644
> --- a/hw/xen/xen_pt_config_init.c
> +++ b/hw/xen/xen_pt_config_init.c
> @@ -575,6 +575,22 @@ static int 
> xen_pt_exp_rom_bar_reg_write(XenPCIPassthroughState *s,
>      return 0;
>  }
>  
> +static int xen_pt_intel_opregion_read(XenPCIPassthroughState *s,
> +                                      XenPTReg *cfg_entry,
> +                                      uint32_t *value, uint32_t valid_mask)
> +{
> +    *value = igd_read_opregion(s);
> +    return 0;
> +}
> +
> +static int xen_pt_intel_opregion_write(XenPCIPassthroughState *s,
> +                                       XenPTReg *cfg_entry, uint32_t *value,
> +                                       uint32_t dev_value, uint32_t 
> valid_mask)
> +{
> +    igd_write_opregion(s, *value);
> +    return 0;
> +}
> +
>  /* Header Type0 reg static information table */
>  static XenPTRegInfo xen_pt_emu_reg_header0[] = {
>      /* Vendor ID reg */
> @@ -1440,6 +1456,20 @@ static XenPTRegInfo xen_pt_emu_reg_msix[] = {
>      },
>  };
>  
> +static XenPTRegInfo xen_pt_emu_reg_igd_opregion[] = {
> +    /* Intel IGFX OpRegion reg */
> +    {
> +        .offset     = 0x0,
> +        .size       = 4,
> +        .init_val   = 0,
> +        .no_wb      = 1,
> +        .u.dw.read   = xen_pt_intel_opregion_read,
> +        .u.dw.write  = xen_pt_intel_opregion_write,
> +    },
> +    {
> +        .size = 0,
> +    },
> +};
>  
>  /****************************
>   * Capabilities
> @@ -1677,6 +1707,14 @@ static const XenPTRegGroupInfo xen_pt_emu_reg_grps[] = 
> {
>          .size_init   = xen_pt_msix_size_init,
>          .emu_regs = xen_pt_emu_reg_msix,
>      },
> +    /* Intel IGD Opregion group */
> +    {
> +        .grp_id      = PCI_INTEL_OPREGION,
> +        .grp_type    = XEN_PT_GRP_TYPE_EMU,
> +        .grp_size    = 0x4,
> +        .size_init   = xen_pt_reg_grp_size_init,
> +        .emu_regs    = xen_pt_emu_reg_igd_opregion,
> +    },
>      {
>          .grp_size = 0,
>      },

If I am not mistaken, in the original patch to qemu-xen-traditional, we
were not adding an Intel IGD Opregion group. Instead we were changing
the size of the header Type0 reg group:

+static uint8_t pt_reg_grp_header0_size_init(struct pt_dev *ptdev,
+        struct pt_reg_grp_info_tbl *grp_reg, uint32_t base_offset)
+{
+    /*
+    ** By default we will trap up to 0x40 in the cfg space.
+    ** If an intel device is pass through we need to trap 0xfc,
+    ** therefore the size should be 0xff.
+    */
+    if (igd_passthru)
+        return 0xFF;
+    return grp_reg->grp_size;
+}

Here instead we are adding the new group and forcing the offset to be
PCI_INTEL_OPREGION, below. It looks functionally equivalent, but nicer.

But wouldn't it be even better to have find_cap_offset return the right
offset for Intel IGD Opregion group? Why do we need to manually set
reg_grp_offset to PCI_INTEL_OPREGION?



> @@ -1806,7 +1844,8 @@ int xen_pt_config_init(XenPCIPassthroughState *s)
>          uint32_t reg_grp_offset = 0;
>          XenPTRegGroup *reg_grp_entry = NULL;
>  
> -        if (xen_pt_emu_reg_grps[i].grp_id != 0xFF) {
> +        if (xen_pt_emu_reg_grps[i].grp_id != 0xFF
> +            && xen_pt_emu_reg_grps[i].grp_id != PCI_INTEL_OPREGION) {
>              if (xen_pt_hide_dev_cap(&s->real_device,
>                                      xen_pt_emu_reg_grps[i].grp_id)) {
>                  continue;
> @@ -1819,6 +1858,10 @@ int xen_pt_config_init(XenPCIPassthroughState *s)
>              }
>          }
>  
> +        if (xen_pt_emu_reg_grps[i].grp_id == PCI_INTEL_OPREGION) {
> +            reg_grp_offset = PCI_INTEL_OPREGION;
> +        }
> +
>          reg_grp_entry = g_new0(XenPTRegGroup, 1);
>          QLIST_INIT(&reg_grp_entry->reg_tbl_list);
>          QLIST_INSERT_HEAD(&s->reg_grps, reg_grp_entry, entries);
> diff --git a/hw/xen/xen_pt_graphics.c b/hw/xen/xen_pt_graphics.c
> index 066bc4d..b25ecae 100644
> --- a/hw/xen/xen_pt_graphics.c
> +++ b/hw/xen/xen_pt_graphics.c
> @@ -6,6 +6,8 @@
>  #include "hw/xen/xen_backend.h"
>  #include "hw/pci/pci_bus.h"
>  
> +static int igd_guest_opregion;
> +
>  static int is_vga_passthrough(XenHostPCIDevice *dev)
>  {
>      return (xen_has_gfx_passthru
> @@ -386,3 +388,48 @@ err_out:
>      XEN_PT_ERR(pci_dev, "Can't get pci_dev_host_bridge\n");
>      return -1;
>  }
> +
> +uint32_t igd_read_opregion(XenPCIPassthroughState *s)
> +{
> +    uint32_t val = 0;
> +
> +    if (igd_guest_opregion == 0) {
> +        return val;
> +    }
> +
> +    val = igd_guest_opregion;
> +
> +    XEN_PT_LOG(&s->dev, "Read opregion val=%x\n", val);
> +    return val;
> +}
> +
> +void igd_write_opregion(XenPCIPassthroughState *s, uint32_t val)
> +{
> +    uint32_t host_opregion = 0;
> +    int ret;
> +
> +    if (igd_guest_opregion) {
> +        XEN_PT_LOG(&s->dev, "opregion register already been set, ignoring 
> %x\n",
> +                   val);
> +        return;
> +    }
> +
> +    xen_host_pci_get_block(&s->real_device, PCI_INTEL_OPREGION,
> +            (uint8_t *)&host_opregion, 4);
> +    igd_guest_opregion = (val & ~0xfff) | (host_opregion & 0xfff);
> +
> +    ret = xc_domain_memory_mapping(xen_xc, xen_domid,
> +            igd_guest_opregion >> XC_PAGE_SHIFT,
> +            host_opregion >> XC_PAGE_SHIFT,
> +            2,
> +            DPCI_ADD_MAPPING);
> +
> +    if (ret != 0) {
> +        XEN_PT_ERR(&s->dev, "Error: Can't map opregion\n");
> +        igd_guest_opregion = 0;
> +        return;
> +    }
> +
> +    XEN_PT_LOG(&s->dev, "Map OpRegion: %x -> %x\n", host_opregion,
> +            igd_guest_opregion);
> +}
> -- 
> 1.9.1
>