On 03/08/2015 08:26 PM, Kevin O'Connor wrote:
On Sun, Mar 08, 2015 at 07:51:42PM +0200, Marcel Apfelbaum wrote:
On 03/08/2015 06:13 PM, Kevin O'Connor wrote:
If I read this correctly, it looks like a machine with two root buses and 20 devices, each with one memory range and one io range, would end up with 40 CRS ranges (ie, a CRS range for every resource).
Correct.
As Michael pointed out in another thread, the firmware is considered guest code and QEMU cannot assume anything on how the resources are assigned. This is why this solution was chosen.
However we have two things that make the situation a little better.
- The PXB implementation includes a pci-bridge and all devices are automatically attached to the secondary bus, in this way we have one IO/MEM range per extra root bus.
Out of curiosity, does the PXB implementation add the pci-bridge just to simplify the IO/MEM range, or are there other technical reasons for it?
We have another elephant there :) -> pci hotplug. All the "free" memory ranges are assigned to bus 0, this will leave the pxb buses without the hotplug capability. Using a PCI bridge will give us some IO/MEM ranges for hotplug: the ones created because of minimum requirement by PCI spec and not used currently by any devices.
On top of this series we can add a merge algorithm that will bring together consecutive ranges. This series does not include this optimization and it focuses on the correctness.
It also
looks like this furthers the requirement that the guest firmware assign the PCI resources prior to QEMU being able to generate the ACPI tables.
Am I correct? If so, that doesn't sound ideal.
You are correct, however is not that bad because we have the following sequence:
- Early in the boot sequence the bios scans the PCI buses and assigns IO/MEM ranges
- At this moment all resources needed by QEMU are present in the configuration space.
- At the end of the boot sequence the BIOS queries the ACPI tables and *only then* the tables are computed.
I think we use that implicitly for other features, anyway, it looks like an elegant solution with no real drawbacks. (Our assumptions are safe)
Thank you for the clarification. I understand that it works, but I've never been that comfortable with the QEMU<->firmware dance with PCI resources. I do understand that the alternatives have as many or more problems though. So, I'm not objecting to this implementation.
No problem, thank you and your review is much appreciated as always, Marcel
Cheers, -Kevin