Re: pmap cleanup approach.

[Cherry G.Mathew <cherry%zyx.in@localhost>]

Manuel Bouyer <bouyer%antioche.eu.org@localhost> writes:

> On Sat, Feb 16, 2019 at 03:23:35PM +0530, Cherry G.Mathew wrote:
>> [...]
>> 
>> Finally, as an experiment for later (once PVHVM is stable) I want to
>> experiment with removing the idea of 'p2m/m2p' tables in their entirety
>> and use the power of our uvm_hotplug(9) infrastructure to directly
>> manage 'machine frames', however fragmented they are.
>
> Not only the memory space may be fragmented, but it's dynamic.
> This means that an in-use segment could be split in 2, (or more likely 3)
> segments. I don't see this functionality in uvm_hotplug(9).
>

uvm_hotplug(9) can manage RAM extents upto page size granularity and do
dynamic insert/remove - what it can't do is to figure out all
dependencies such as mapping/swap related updates.

Does xen pull the rug from underneath us with the p2m/m2p tables ? ie;
would the tables be updated by the hypervisor when this dynamic action
happens rather than via a fault or other callback ?

> Also I fear that managing m2p updates via UVM makes these operations
> much heavier.

Yeah, that's one thing. Nevertheless I'd like to give it a spin to see
how it looks like - we could make some performance comparisons at the
very least.

-- 
~cherry