[Qemu-devel] Re: [SeaBIOS] seabios: acpi: add _RMV control method for PCI devices

[Gleb Natapov <gleb@redhat.com>]

On Mon, Dec 13, 2010 at 10:53:07AM +0200, Avi Kivity wrote:
> On 12/13/2010 10:49 AM, Gleb Natapov wrote:
> >On Mon, Dec 13, 2010 at 10:41:25AM +0200, Avi Kivity wrote:
> >>  On 12/13/2010 02:00 AM, Marcelo Tosatti wrote:
> >>  >On Sat, Dec 11, 2010 at 09:39:30AM +0200, Avi Kivity wrote:
> >>  >>   On 12/08/2010 07:08 PM, Marcelo Tosatti wrote:
> >>  >>   >Use _RMV method to indicate whether device can be removed.
> >>  >>   >
> >>  >>   >Data is retrieved from QEMU via I/O port 0xae0c.
> >>  >>   >
> >>  >>
> >>  >>   Where did this port come from?
> >>  >
> >>  >Its the next available address after "PCI EJ base", used
> >>  >for QEMU<->ACPI hotplug communication.
> >>  >
> >>  >>   What's the protocol?
> >>  >
> >>  >ACPI reads the 32-bit field indicating the return value of the _RMV
> >>  >method (which is used by Windows to decide removability). 1-bit per
> >>  >slot.
> >>  >
> >>  >More ports have to be registered if more buses are added.
> >>  >
> >>  >>   Maybe we should do this via fw_cfg.
> >>  >
> >>  >I don't see a need for it? (yes, it might be possible, but i'm not
> >>  >familiar enough with AML).
> >>
> >>  To avoid adding tons of undocumented I/O ports, and to allow
> >>  discoverability (what happens with a new seabios on old qemu)?
> >>
> >We already have out own mini pci hot-plug controller at io port 0xae00.
> >The patch just extends its functionality a bit. Logically this
> >functionality belongs there.
> 
> Well, at least it should be documented.
> 
Agree.

> We could also deprecate the old port and use fwcfg for everything
> (try fwcfg, fall back to ae00).
> 
fwcfg designed to be simple for easy use by firmware. It has two port
one for index another for value, so its use is racy in multi-threaded SMP
environment. DSDT code is executed in such environment. There is lock
facility in AML, but why complicate things.
 
> >>  We could do this in two ways: by adding a fwcfg client to the DSDT,
> >>  or by copying the information to system memory, and referencing
> >>  system memory from the DSDT.
> >>
> >This is even worse. It requires some fixed address to be shared between
> >DSDT and Seabios (or alternatively Seabios will have to generate this
> >part of DSDT dynamically).
> >
> 
> Could easily be something in the F segment.
> 
Yes, but then we will have two magic values (fwcfg index + address
in F segment) instead of one (address of pci hot-plug controller).

--
			Gleb.