Re: [Qemu-devel] [PATCH 11/19] target-i386: introduce apic-id property

[Eduardo Habkost]

On Mon, Apr 15, 2013 at 10:27:57PM +0200, Igor Mammedov wrote:
> On Mon, 15 Apr 2013 11:49:40 -0300
> Eduardo Habkost <address@hidden> wrote:
> 
> > On Mon, Apr 15, 2013 at 04:34:41PM +0200, Igor Mammedov wrote:
> > > On Mon, 15 Apr 2013 10:45:24 -0300
> > > Eduardo Habkost <address@hidden> wrote:
> > > 
> > > > On Mon, Apr 15, 2013 at 03:37:00PM +0200, Igor Mammedov wrote:
> > > > [...]
> > > > > > > > ID to be directly specified in the device_add/-device options.
> > > > > > > > That's how real CPUs work: as the CPU manufacturer doesn't know
> > > > > > > > what will be the package ID of the CPU, the APIC IDs are not
> > > > > > > > hardcoded in the CPU; they are calculated based on the CPU 
> > > > > > > > topology
> > > > > > > > and some socket identifier signal coming from the board.
> > > > > > > that is why apic_id has been made a property, to be set from 
> > > > > > > outside.
> > > > > > 
> > > > > > True. I believe the conflict here is: we want other objects to set 
> > > > > > the
> > > > > > APIC ID (be it the board, or socket/core objects), but at the same 
> > > > > > time
> > > > > > it would be interesting to not expose the APIC ID outside QEMU. 
> > > > > > Being
> > > > > > too flexible regarding the APIC ID is more likely to cause problems
> > > > > > later.
> > > > > > 
> > > > > > That said, I don't mind having a "apic-id" property because it is 
> > > > > > easier
> > > > > > to simply expose it directly. But do you agree that: 1) we don't 
> > > > > > really
> > > > > > need to expose it to be set from outside QEMU; 2) we shouldn't 
> > > > > > require
> > > > > > it to be set from outside QEMU; 3) we should recommend users to not 
> > > > > > try
> > > > > > to fiddle it with?
> > > > > Due to nature of per thread CPU hotplug, management will have to 
> > > > > specify
> > > > > some kind of ID to specify which CPU is being plugged. Management 
> > > > > really
> > > > > doesn't/shouldn't care what this ID is.
> > > > 
> > > > As long as management really doesn't/shouldn't care what the ID is,
> > > > exposing the APIC ID in the form of an opaque CPU identifier wouldn't be
> > > > a problem to me. I just wanted to make clarify if we agree that messing
> > > > with the APIC ID directly won't be recommended and that the "apic-id"
> > > > property will be for QEMU internal use only.
> > > On contrary, it's useful external feature, x86 guests see only APIC ID, 
> > > since
> > > it's the only ID they [should] know about. So guest aware mgmt could
> > > definitely use apic_id propery to correlate CPU in guest with QEMU view of
> > > them.
> > 
> > You're right, _reading_ the APIC ID is very useful. I am worried about
> > _setting_ it from external code.
> 
> currently it's not possible since cpu-add doesn't allow to set any properties.
> 
> We will need setting it for device_add though.

Not necessarily. That's why I am insisting on an interface based on
links/topology, not based on a raw "apic-id" property: instead of
setting apic-id directly, we could just require that the CPU be attached
to the right socket/core objects, and the APIC ID would be magically
calculated correctly.

Or we could just let the right socket/core/thread IDs to be set as
properties, and apic-id could be calculated based on that. There are
many ways to expose an abstraction that's simpler to use and less likely
to cause problems.


> By then, I guess some way to check that it's valid would be enough, otherwise
> hot-plugged CPU will be out of scope of MADT and guest would ignore it or
> through an error.
> 

-- 
Eduardo