On Thu, Dec 24, 2009 at 12:22:12PM -0500, Kevin O'Connor wrote:
On Thu, Dec 24, 2009 at 06:54:55PM +0200, Gleb Natapov wrote:
On Thu, Dec 24, 2009 at 11:38:45AM -0500, Kevin O'Connor wrote:
foreachpci(bdf, max) {
if (pci_bdf_to_bus(bdf) > 0)break;mptable routing info should be created for each pci bus.
My read of the spec is that independent PCI buses need to be added to the MPTable, but a bus created due to a PCI-to-PCI bridge (eg, on an add-on card) need not go into the table.
See example here:
Ah so pci_bdf_to_bus(bdf) > 0 for the bus behind PCI-to-PCI bridge? I don't wee why you read spec this way. Can you point me to the part of the spec that makes you think so?
In that example, it looks like the machine has two independent buses - note how bus 1 has more than four irqs - so it can't be an add-on card.
Yes. In this example there are several PCI domains.
Hmm. It looks like we need to create io_bus entries dynamically for each newly discovered PCI bus.
Right now, the pciinit.c code handles neither PCI-to-PCI bridges nor independent PCI bridges. I think it would be useful to support bridges (for mapping in add-on cards that have them), but I don't know why an emulator would want to have independent buses. Note that it's difficult to detect an independent bus - some sideband mechanism is needed (eg, see CONFIG_PCI_ROOT1 in src/config.h).
-- Gleb.