Re: [PATCH] Setup vcpu add/remove infrastructure, including madt bios_info and dsdt.

[Gleb Natapov <gleb@redhat.com>]

On Fri, Jan 22, 2010 at 06:54:42PM +0800, Liu, Jinsong wrote:
> Gleb Natapov wrote:
> > On Fri, Jan 22, 2010 at 05:08:32PM +0800, Liu, Jinsong wrote:
> >> Gleb Natapov wrote:
> >>> On Fri, Jan 22, 2010 at 10:15:44AM +0800, Liu, Jinsong wrote:
> >>>> Gleb Natapov wrote:
> >>>>> On Thu, Jan 21, 2010 at 07:48:23PM +0800, Liu, Jinsong wrote:
> >>>>>>> From cb997030cba02e7e74a29b3d942aeba9808ed293 Mon Sep 17
> >>>>>>> 00:00:00 2001
> >>>>>> From: Liu, Jinsong <jinsong.liu@intel.com>
> >>>>>> Date: Fri, 22 Jan 2010 03:18:46 +0800
> >>>>>> Subject: [PATCH]         Setup vcpu add/remove infrastructure,
> >>>>>> including madt bios_info and dsdt.
> >>>>>> 
> >>>>>>         1. setup madt bios_info structure, so that static dsdt
> >>>>>>            get run-time madt info like checksum address, lapic
> >>>>>>            address, max cpu numbers, with least hardcode magic
> >>>>>>            number (realmode address of bios_info).
> >>>>>>         2. setup vcpu add/remove dsdt infrastructure, including
> >>>>>>            processor related acpi objects and control methods.
> >>>>>>            vcpu add/remove will trigger SCI and then control
> >>>>>>            method _L02. By matching madt, vcpu number and
> >>>>>>            add/remove action were found, then by notify control
> >>>>>> method, it will notify OS acpi driver. 
> >>>>>> 
> >>>>>>         Signed-off-by: Liu, Jinsong <jinsong.liu@intel.com>
> >>>>> It looks like AML code is a port of what we had in BOCHS bios with
> >>>>> minor changes. Can you detail what is changed and why for easy
> >>>>> review please? And this still doesn't work with Windows I assume.
> >>>>> 
> >>>> 
> >>>> Yes, my work is based on BOCHS infrastructure, thanks BOCHS :)
> >>>> I just change some minor points:
> >>>> 1. explicitly define returen value of '_MAT' as 'buffer', otherwise
> >>>> some linux acpi driver (i.e. linux 2.6.30) would parse error which
> >>>> will handle it as 'integer' not 'buffer';
> >>>> 2. keep correct 'checksum' of madt when vcpu add/remove, otherwise
> >>>> it will report 'checksum error' when using acpi tools to get madt
> >>>> info if we add/remove vcpu; 
> >>>> 3. add '_EJ0' so that linux has acpi obj under
> >>>> /sys/devices/LNXSYSTM:00, which is need for vcpu remove;
> >>>> 4. on Method(PRSC, 0), just scan 'xxx' vcpus that qemu get from
> >>>> cmdline para 'maxcpus=xxx', not all 256 vcpus, otherwise under some
> >>>> dsdt processor define, it will result error;
> >>> What kind of errors? Qemu should never set bit over maxcpus in PRS.
> >>> 
> >> suppose cmdline define vcpus=4, maxvcpus=8
> >> in original BOCHS, will scan 15 lapic items start from lapic0 of
> >> madt, where it only has maxvcpus lapic items in madt table, hence
> >> there is risk to scan over boundary, scan to other acpi table, and
> >> result in wrong vcpu online/offline status (in my test, I meet this
> >> situation). Because of this reason, this patch limit scan maxvcpus
> >> lapic of madt.    
> >> 
> > But what if cmdline will use vcpu=64? The idea was that \_PR scope
> > will reside in its own ssdt and for each maxvcpus value there will be
> > ssdt with exactly this number of processors. Ideally ssdt will be
> > created dynamically like it is done now, but another solution is to
> > create them at bios compilation time and load correct one at runtime.
> > 
> 
> It's OK for vcpu=64. vcpu<maxvcpus.
> if maxvcpus > processor defined in dsdt, it's OK since no risk scan (bios only support 15 processor is another story);
> if processor defined in dsdt > maxvcpus, it has risk to scan over boundary.
> 
> 
Yes, correct. So why not export maxcpus to DSDT and at the beginning of
NTFY check that Arg0 < maxcpus then?

> > BTW seabios creates very simple ssdt with \_PR scope right now and I
> > don't see yous patch removes this so you end up with two conflicting 
> > \_PR scopes.
> 
> Yes, I just notice \_PR at acpi.c.
> we can move \_PR to ssdt after ssdt infrastructure was full setup.
> 
> Thanks,
> Jinsong

--
			Gleb.