Re: [Qemu-stable] [Qemu-devel] [PATCH] intel_iommu: handle invalid ce for shadow sync

[Jason Wang]

On 2018年09月13日 15:55, Peter Xu wrote:
> There are two callers for vtd_sync_shadow_page_table_range(), one
> provided a valid context entry and one not.  Move that fetching
> operation into the caller vtd_sync_shadow_page_table() where we need to
> fetch the context entry.
>
> Meanwhile, we should handle VTD_FR_CONTEXT_ENTRY_P properly when
> synchronizing shadow page tables.  Having invalid context entry there is
> perfectly valid when we move a device out of an existing domain.  When
> that happens, instead of posting an error we invalidate the whole region.
>
> Without this patch, QEMU will crash if we do these steps:
>
> (1) start QEMU with VT-d IOMMU and two 10G NICs (ixgbe)
> (2) bind the NICs with vfio-pci in the guest
> (3) start testpmd with the NICs applied
> (4) stop testpmd
> (5) rebind the NIC back to ixgbe kernel driver
>
> The patch should fix it.
>
> Reported-by: Pei Zhang <address@hidden>
> Tested-by: Pei Zhang <address@hidden>
> CC: Pei Zhang <address@hidden>
> CC: Alex Williamson <address@hidden>
> CC: Jason Wang <address@hidden>
> CC: Maxime Coquelin <address@hidden>
> CC: Michael S. Tsirkin <address@hidden>
> CC: QEMU Stable <address@hidden>
> Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1627272
> Signed-off-by: Peter Xu <address@hidden>
> ---
>   hw/i386/intel_iommu.c | 54 ++++++++++++++++++++++++++-----------------
>   1 file changed, 33 insertions(+), 21 deletions(-)

Reviewed-by: Jason Wang <address@hidden>

Some nits, see below.

> diff --git a/hw/i386/intel_iommu.c b/hw/i386/intel_iommu.c
> index 3dfada19a6..2509520d6f 100644
> --- a/hw/i386/intel_iommu.c
> +++ b/hw/i386/intel_iommu.c
> @@ -37,6 +37,8 @@
>   #include "kvm_i386.h"
>   #include "trace.h"
>   
> +static void vtd_address_space_unmap(VTDAddressSpace *as, IOMMUNotifier *n);
> +
>   static void vtd_define_quad(IntelIOMMUState *s, hwaddr addr, uint64_t val,
>                               uint64_t wmask, uint64_t w1cmask)
>   {
> @@ -1047,39 +1049,49 @@ static int 
> vtd_sync_shadow_page_table_range(VTDAddressSpace *vtd_as,
>           .notify_unmap = true,
>           .aw = s->aw_bits,
>           .as = vtd_as,
> +        .domain_id = VTD_CONTEXT_ENTRY_DID(ce->hi),
>       };
> -    VTDContextEntry ce_cache;
> +
> +    return vtd_page_walk(ce, addr, addr + size, &info);
> +}
> +
> +static int vtd_sync_shadow_page_table(VTDAddressSpace *vtd_as)
> +{
>       int ret;
> +    VTDContextEntry ce;
> +    IOMMUNotifier *n;
>   
> -    if (ce) {
> -        /* If the caller provided context entry, use it */
> -        ce_cache = *ce;
> -    } else {
> -        /* If the caller didn't provide ce, try to fetch */
> -        ret = vtd_dev_to_context_entry(s, pci_bus_num(vtd_as->bus),
> -                                       vtd_as->devfn, &ce_cache);
> -        if (ret) {
> +    ret = vtd_dev_to_context_entry(vtd_as->iommu_state,
> +                                   pci_bus_num(vtd_as->bus),
> +                                   vtd_as->devfn, &ce);
> +    if (ret) {
> +        if (ret == -VTD_FR_CONTEXT_ENTRY_P) {
> +            /*
> +             * It's a valid scenario to have a context entry that is
> +             * not present.  For example, when a device is removed
> +             * from an existing domain then the context entry will be
> +             * zeroed by the guest before it was put into another
> +             * domain.  When this happens, instead of synchronizing
> +             * the shadow pages we should invalidate all existing
> +             * mappings and notify the backends.
> +             */
> +            IOMMU_NOTIFIER_FOREACH(n, &vtd_as->iommu) {
> +                vtd_address_space_unmap(vtd_as, n);
> +            }
> +        } else {
>               /*
>                * This should not really happen, but in case it happens,
>                * we just skip the sync for this time.  After all we even
>                * don't have the root table pointer!
>                */

It looks to me the comment is not accurate, no root pointer is not the 
only reason for the failure of vtd_dev_to_context_entry().

>               error_report_once("%s: invalid context entry for bus 0x%x"
> -                              " devfn 0x%x",
> -                              __func__, pci_bus_num(vtd_as->bus),
> -                              vtd_as->devfn);
> -            return 0;

I'm not quite sure error_report_once() is really needed here since all 
failures has been traced.

> +                              " devfn 0x%x", __func__,
> +                              pci_bus_num(vtd_as->bus), vtd_as->devfn);
>           }
> +        return 0;
>       }
>   
> -    info.domain_id = VTD_CONTEXT_ENTRY_DID(ce_cache.hi);
> -
> -    return vtd_page_walk(&ce_cache, addr, addr + size, &info);
> -}
> -
> -static int vtd_sync_shadow_page_table(VTDAddressSpace *vtd_as)
> -{
> -    return vtd_sync_shadow_page_table_range(vtd_as, NULL, 0, UINT64_MAX);
> +    return vtd_sync_shadow_page_table_range(vtd_as, &ce, 0, UINT64_MAX);
>   }

As has been discussed, this will left addr UINT64_MAX, it's better to 
have [start, end] instead of (start, range).

Thanks

>   
>   /*