Re: [PATCH v3 19/49] kvm: Make kvm_convert_memory() obey ram_block_discard_is_enabled()

[Paolo Bonzini <pbonzini@redhat.com>]

On 3/20/24 09:39, Michael Roth wrote:
> Some subsystems like VFIO might disable ram block discard for
> uncoordinated cases. Since kvm_convert_memory()/guest_memfd don't
> implement a RamDiscardManager handler to convey discard operations to
> various listeners like VFIO. > Because of this, sequences like the
> following can result due to stale IOMMU mappings:

Alternatively, should guest-memfd memory regions call 
ram_block_discard_require(true)?  This will prevent VFIO from operating, 
but it will avoid consuming twice the memory.

If desirable, guest-memfd support can be changed to implement an 
extension of RamDiscardManager that notifies about private/shared memory 
changes, and then guest-memfd would be able to support coordinated 
discard.  But I wonder if that's doable at all - how common are 
shared<->private flips, and is it feasible to change the IOMMU page 
tables every time?

If the real solution is SEV-TIO (which means essentially guest_memfd 
support for VFIO), calling ram_block_discard_require(true) may be the 
simplest stopgap solution.

Paolo

>    - convert page shared->private
>    - discard shared page
>    - convert page private->shared
>    - new page is allocated
>    - issue DMA operations against that shared page
> 
> Address this by taking ram_block_discard_is_enabled() into account when
> deciding whether or not to discard pages.
> 
> Signed-off-by: Michael Roth <michael.roth@amd.com>
> ---
>   accel/kvm/kvm-all.c | 8 ++++++--
>   1 file changed, 6 insertions(+), 2 deletions(-)
> 
> diff --git a/accel/kvm/kvm-all.c b/accel/kvm/kvm-all.c
> index 53ce4f091e..6ae03c880f 100644
> --- a/accel/kvm/kvm-all.c
> +++ b/accel/kvm/kvm-all.c
> @@ -2962,10 +2962,14 @@ static int kvm_convert_memory(hwaddr start, hwaddr size, bool to_private)
>                   */
>                   return 0;
>               } else {
> -                ret = ram_block_discard_range(rb, offset, size);
> +                ret = ram_block_discard_is_disabled()
> +                      ? ram_block_discard_range(rb, offset, size)
> +                      : 0;
>               }
>           } else {
> -            ret = ram_block_discard_guest_memfd_range(rb, offset, size);
> +            ret = ram_block_discard_is_disabled()
> +                  ? ram_block_discard_guest_memfd_range(rb, offset, size)
> +                  : 0;
>           }
>       } else {
>           error_report("Convert non guest_memfd backed memory region "