[Qemu-devel] Re: questions on default_config_write in hw/pci.c

[Marcelo Tosatti <mtosatti@redhat.com>]

On Thu, Apr 30, 2009 at 08:29:55PM +0300, Michael S. Tsirkin wrote:
> On Thu, Apr 30, 2009 at 02:15:26PM -0300, Marcelo Tosatti wrote:
> > On Thu, Apr 30, 2009 at 07:15:08PM +0300, Michael S. Tsirkin wrote:
> > > Hi,
> > > I've been looking at hw/pci.c, specifically at how config
> > > read/write are implemented, and have a couple of questions
> > > about default_config_write:
> > > 
> > > 1. The code at the beginning (if len == 4 ...)
> > >    seems to only update pci base registers if a dword write 
> > >    is performed. I think it's legal for the guest to perform 4
> > >    single-byte writes. Should this be supported?
> > > 
> > > 2. The large switch statement at the end of this function
> > >    uses hard-coded register offsets. Would it make sense
> > >    to change it to use macros from hw/pci.h?
> > > 
> > > 3. Still there I see:
> > >         switch(d->config[0x0e]) {
> > >         case 0x00:
> > >         case 0x80:
> > >    register 0x0e is header type, which has defined values
> > >    of 0x00 (device or host bridge), 0x01 (pci to pci bridge) and
> > >    0x02 (cardbus bridge). What is 0x80 and when is it used?
> > >    Would it make sense to remove this?
> > 
> > Don't know. Check the PCI spec?
> 
> That's the first thing I did :) What I site above is from there:
> "The Header Type field (located at offset 0Eh) defines what layout is
> provided. Currently three Header Types are defined, 00h which has the
> layout shown in Figure 6-1, 01h which is defined for PCI-to-PCI bridges
> and is documented in the PCI to PCI Bridge Architecture Specification,
> and 02h which is defined for CardBus bridges and is documented in the PC
> Card Standard."
> 
> > > 4. Still there, there's some handling done for type 1 devices.
> > >    This support seems imcomplete.
> > >    Are there any PCI-to-PCI bridges emulated by qemu?
> > >    Would it make sense to remove this code?
> > 
> > 
> > It did work at one point:
> > 
> > http://www.mail-archive.com/kvm-devel@lists.sourceforge.net/msg16647.html
> 
> Sigh. I guess I'll have to fix it rather than just rip it out then ...
> What I refer to first of all is that writes to BAR registers go through
> logic for regular PCI devices, which have a different layout.
> Does this make sense for some reason?

The first 16 bytes of the config space are common (which include
vendor/device). But some refactoring to have the header specific part
into separate functions would be nice?

Anthony: it did work for me, all 128 slots.