Re: [PATCH] hw/i386/pc: fix max_used_gpa for 32-bit systems

[Ani Sinha]

On Tue, Sep 19, 2023 at 9:53 AM Ani Sinha <anisinha@redhat.com> wrote:
>
> On Tue, Sep 19, 2023 at 9:20 AM Ani Sinha <anisinha@redhat.com> wrote:
> >
> > On Mon, Sep 18, 2023 at 9:28 PM David Hildenbrand <david@redhat.com> wrote:
> > >
> > > On 18.09.23 17:56, Ani Sinha wrote:
> > > > On Mon, Sep 18, 2023 at 8:59 PM David Hildenbrand <david@redhat.com> 
> > > > wrote:
> > > >>
> > > >> On 18.09.23 17:22, Ani Sinha wrote:
> > > >>> On Mon, Sep 18, 2023 at 7:25 PM Ani Sinha <anisinha@redhat.com> wrote:
> > > >>>>
> > > >>>> 32-bit systems do not have a reserved memory for hole64 but they may 
> > > >>>> have a
> > > >>>> reserved memory space for memory hotplug. Since, hole64 starts after 
> > > >>>> the
> > > >>>> reserved hotplug memory, the unaligned hole64 start address gives us 
> > > >>>> the
> > > >>>> end address for this memory hotplug region that the processor may 
> > > >>>> use.
> > > >>>> Fix this. This ensures that the physical address space bound 
> > > >>>> checking works
> > > >>>> correctly for 32-bit systems as well.
> > > >>>
> > > >>> This patch breaks some unit tests. I am not sure why it did not catch
> > > >>> it when I tested it before sending.
> > > >>> Will have to resend after fixing the tests.
> > > >>
> > > >> Probably because they supply more memory than the system can actually
> > > >> handle? (e.g., -m 4g on 32bit)?
> > > >
> > > > cxl tests are failing for example.
> > > >
> > > > $ ./qemu-system-i386 -display none -machine q35,cxl=on
> > > > qemu-system-i386: Address space limit 0xffffffff < 0x1000fffff
> > > > phys-bits too low (32)
> >
> > also another thing is:
> >
> > ./qemu-system-i386 -machine pc -m 128
> > works but ...
> >
> > $ ./qemu-system-i386 -machine pc -m 128,slots=3,maxmem=1G
> > qemu-system-i386: Address space limit 0xffffffff < 0x1f7ffffff
> > phys-bits too low (32)
> >
> > or
> >
> > $ ./qemu-system-i386 -machine pc-i440fx-8.2 -accel kvm -m 
> > 128,slots=3,maxmem=1G
> > qemu-system-i386: Address space limit 0xffffffff < 0x1f7ffffff
> > phys-bits too low (32)
> >
> > but of course after the compat knob older pc machines work fine using
> > the old logic :
> >
> > $ ./qemu-system-i386 -machine pc-i440fx-8.1 -accel kvm -m 
> > 128,slots=3,maxmem=1G
> > VNC server running on ::1:5900
> > ^Cqemu-system-i386: terminating on signal 2
>
> I dpn't know if we always need to do this but this code adds 1 GiB per
> slot for device memory :
>
>     if (pcmc->enforce_aligned_dimm) {
>          /* size device region assuming 1G page max alignment per slot */
>          size += (1 * GiB) * machine->ram_slots;
>      }
>
> For a 32-bit machine that is a lot of memory consumed in just alignment.

Let's look at an example when we get rid of all alignment stuff.

$ ./qemu-system-i386 -machine pc-i440fx-8.2 -m 512M,slots=1,maxmem=1G
above 4G start: 0x100000000,above 4G size: 0x0
qemu-system-i386: Address space limit 0xffffffff < 0x11ffffffe
phys-bits too low (32)

So basically, above_4g_start = 4GiB. size = 0.
Then it is adding the device memory which is 1GiB - 0.5 GiB = 0.5 GiB.
So the  0x11ffffffe is exactly 4.5 GiB.

Anything above 4 GiB is beyond 32 bits.