Yes, this will reduce BQL context in lockless IO a lot. And I'll based
on your patch to extract HPETTimer from BqlRefCell.

> Preserving the old migration format can then be solved in two ways:
>
> 1) with a handwritten ToMigrationStateShared implementation for
> HPETTimer (and marking the tn_regs array as #[migration_state(omit)])

Yes, compared with 2), I also this is the better choice, which looks
more common and could be an good example for other device.

> 2) by also adding num_timers_save and the timer's expiration to
> HPETRegisters and HPETTimerRegisters, respectively.
>
> I think I prefer the former, but I haven't written the code so it's
> not my choice. :)

> @@ -181,6 +181,9 @@ fn timer_handler(timer_cell: &BqlRefCell<HPETTimer>) {
>  #[repr(C)]
>  #[derive(Debug, Default, ToMigrationState)]
>  pub struct HPETTimerRegisters {
> +    // Only needed for migration
> +    index: u8,

I didn't get the point why we need to migrate this "index" instead of
HPETTimer::index?

Anyway, if we continue to keep index in HPETTimerRegisters, we can make
it have usize type with a "#[migration_state(try_into(u8))]" property.

If we just HPETTimer::index for migration, I think it's possible to do
type convertion in handwritten ToMigrationStateShared.

Thanks,
Zhao