On Wed, Jun 08, 2011 at 11:11:37PM +0200, Jan Kiszka wrote: > On 2011-06-08 23:09, Michael S. Tsirkin wrote: > > On Wed, Jun 08, 2011 at 11:02:44PM +0200, Jan Kiszka wrote: > >> On 2011-06-08 23:00, Michael S. Tsirkin wrote: > >>> On Wed, Jun 08, 2011 at 10:48:10PM +0200, Jan Kiszka wrote: > >>>> On 2011-06-08 21:53, Michael S. Tsirkin wrote: > >>>>> On Wed, Jun 08, 2011 at 06:21:51PM +0200, Jan Kiszka wrote: > >>>>>> Add PCI_MSIX_TABLE and PCI_MSIX_PBA, align other MSIX related constant > >>>>>> names to libpci style. Will be used for device assignment code in > >>>>>> qemu-kvm. > >>>>>> > >>>>>> Signed-off-by: Jan Kiszka <jan.kis...@siemens.com> > >>>>> > >>>>> Besides keeping pci_regs.h aligned with the original, > >>>>> I also think ideally pci register banging should stay > >>>>> within the pci subsystem. > >>>>> > >>>>> Could we add high-level APIs to help with that, > >>>>> instead of having kvm look at config space directly? > >>>> > >>>> We could move the related static inlines from msi/msix.c to the headers > >>>> in order to test for bits etc. Still, kvm needs to interpret the config > >>>> space of the assigned device, so the abstraction will remain rather low. > >>>> > >>>> Jan > >>>> > >>> > >>> Hmm, at least for MSI/MSIX I thought this is done by kvm in kernel? > >>> > >> > >> At least for the "traditional" assignment interface (VFIO may offload > >> something), no. User space does the cap analysis, filtering, and in the > >> MSI/MSI-X case the translation to QEMU msi/msix services. The latter is > >> even WIP in my tree. Surrent assignment open-codes this, missing many > >> corner cases. > >> > >> Jan > >> > > > > Anyway, if some defines need to be in a header, and aren't upstream > > yet, let's create pci_ext_regs.h and add a comment there that we > > should work on upstreaming them. > > Sounds good. But what is supposed to be upstream for us, the kernel or > pci-utils/libpci? > > Jan >
It's currently the kernel, I don't really see a reason to change that. -- MST