From: Hans-Jörg Höxer <hshoexer@genua.de>
Subject: Re: [EXT] Reworking VMM's nested paging & guest memory (de-vmspace-ification)
To: <tech@openbsd.org>
Date: Mon, 28 Apr 2025 10:50:44 +0200

Hi,


On Fri, Apr 25, 2025 at 01:05:27PM -0400, Dave Voutila wrote:
>...
> >> @@ -487,13 +527,13 @@ vm_create_check_mem_ranges(struct vm_create_params *vc
> >>  		 * Calling uvm_share() when creating the VM will take care of
> >>  		 * further checks.
> >>  		 */
> >> -		if (vmr->vmr_va < VM_MIN_ADDRESS ||
> >> +/*		if (vmr->vmr_va < VM_MIN_ADDRESS ||
> >>  		    vmr->vmr_va >= VM_MAXUSER_ADDRESS ||
> >>  		    vmr->vmr_size >= VM_MAXUSER_ADDRESS - vmr->vmr_va) {
> >>  			DPRINTF("guest va not within range or wraps\n");
> >>  			return (0);
> >>  		}
> >> -
> >> +*/
> >
> > Why remove this check?  Shouldn't these constraints still hold?
> >
> 
> I'm shifting to a model where the userland process no longer dictates
> the virtual address layout of the guest memory in the vmd process. The
> vm_create() and vm_share_mem() functions should be determining the va
> ranges in the vmd proess via uvm_map and now communicating them to the
> calling process.
> 
> In other words, it's turning into an output from the ioctl and not an
> input.
> 
> I'll be sending out a revised diff now that I've iterated on the uvm
> portion with mpi that should make it clear.

I see, thanks!