Re: [PATCH v2 1/2] tools/libs/guest: re-read dominfo after changing a PV guest's width on restore

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 11/11/2026 5:14 pm, George Dunlap wrote:
> The x86 PV restore code reads the domain's information once, when the
> stream starts, and keeps it in ctx->dominfo.  When the X86_PV_INFO
> record gives a width other than the domain's, handle_x86_pv_info()
> changes it with XEN_DOMCTL_set_address_size, and then refreshes only
> the guest's width and page-table levels.
>
> Since 8a9255405ca6 ("x86/domain: adjust limitation on shared_info
> allocation below 4G"), making a domain 32-bit may replace a
> shared_info page (if the originally allocated page was above 4GiB).
> The shared_info_frame read when the stream started will refer to a
> freed page.  handle_shared_info() maps that frame to write the
> SHARED_INFO record, and process_start_info() puts it in the guest's
> start_info.  Xen holds no reference to the replaced page by then, so
> the mapping fails, and with it the restore of any 32-bit PV guest
> whose shared_info page was moved.
>
> Re-read the domain's information after changing its width.
>
> Fixes: 8a9255405ca6 ("x86/domain: adjust limitation on shared_info allocation 
> below 4G")
> Assisted-by: Claude Code:claude-opus-5-5
> Signed-off-by: George Dunlap <gwd@xxxxxxxxxxxxxx>
> ---
> Not tested end to end: no 32-bit PV guest here can suspend.  Restore's
> steps replayed from dom0 on staging, on an Intel box (create a PV
> domain, read its info, make it 32-bit, populate 1GiB, then map the
> frame read first): that map fails with -EBUSY every time, while the
> frame read after the switch maps.  Build-tested.
> ---
>  tools/libs/guest/xg_sr_restore_x86_pv.c | 6 ++++++
>  1 file changed, 6 insertions(+)
>
> diff --git a/tools/libs/guest/xg_sr_restore_x86_pv.c 
> b/tools/libs/guest/xg_sr_restore_x86_pv.c
> index 876748c11e..b72cacd000 100644
> --- a/tools/libs/guest/xg_sr_restore_x86_pv.c
> +++ b/tools/libs/guest/xg_sr_restore_x86_pv.c
> @@ -646,6 +646,12 @@ static int handle_x86_pv_info(struct xc_sr_context *ctx,
>          }
>  
>          /* Domain's information changed, better to refresh. */
> +        if ( xc_domain_getinfo_single(xch, ctx->domid, &ctx->dominfo) < 0 )
> +        {
> +            PERROR("Failed to refresh dominfo for d%d", ctx->domid);
> +            return -1;
> +        }
> +
>          rc = x86_pv_domain_info(ctx);
>          if ( rc != 0 )
>          {

Oh lovely.

I think I'm going to start by putting together a real migration test
including PV32.   We *need* to stop errors this simple from slipping in,
and there have been quite a few recently.

~Andrew