On 03/06/2015 16:05, Igor Mammedov wrote:
>>> > > + rsvd_hva = memory_region_find_rsvd_hva(section->mr);
>>> > > + if (rsvd_hva.mr) {
>>> > > + start_addr = rsvd_hva.offset_within_address_space;
>>> > > + size = int128_get64(rsvd_hva.size);
>>> > > + ram = memory_region_get_ram_ptr(rsvd_hva.mr);
>>> > > + } else {
>>> > > + ram = memory_region_get_ram_ptr(section->mr) +
>>> > > section->offset_within_region;
>>> > > + }
>> >
>> > I don't think this is needed.
>> >
>> > What _could_ be useful is to merge adjacent ranges even if they are
>> > partly unmapped, but your patch doesn't do that.
> merging/splitting for adjacent regions is done at following
> vhost_dev_(un)assign_memory() but it doesn't cover cases with
> gaps in between.
>
> Trying to make merging/splitting work with gaps might be more
> complicated (I haven't tried though), than just passing known
> in advance whole rsvd_hva range.
>
> More over if/when initial memory also converted to rsvd_hva
> (aliasing stopped me there for now), we could throw away all
> this merging and just keep a single rsvd_hva range for all RAM here.
Understood now. This still should be a separate patch. I'm much more
confident with the other two (e.g. what happens if a malicious guest
writes to memory that is still MAP_NORESERVE), so feel free to post
those without RFC tag. But the vhost one really needs mst's eyes.
Paolo
