Bug 1410578 - PCIe: Add an option to PCIe ports to disable IO port space support
Summary: PCIe: Add an option to PCIe ports to disable IO port space support
Status: CLOSED INSUFFICIENT_DATA
Alias: None
Product: Red Hat OpenStack
Classification: Red Hat
Component: openstack-nova
Version: unspecified
Hardware: Unspecified
OS: Unspecified
low
low
Target Milestone: ---
: ---
Assignee: nova-maint
QA Contact: nova-maint
URL:
Whiteboard:
Keywords: FutureFeature, Triaged
Depends On: 1408810 1344299
Blocks: 1410577
TreeView+ depends on / blocked
 
Reported: 2017-01-05 19:27 UTC by Marcel Apfelbaum
Modified: 2018-08-28 09:49 UTC (History)
28 users (show)

(edit)
Clone Of: 1408810
(edit)
Last Closed: 2018-08-28 09:49:19 UTC


Attachments (Terms of Use)

Description Marcel Apfelbaum 2017-01-05 19:27:00 UTC
+++ This bug was initially created as a clone of Bug #1408810 +++

+++ This bug was initially created as a clone of Bug #1344299 +++

Even if the firmware skips assigning IO ranges to PCIe ports (root
ports/downstream ports), Linux guests will still try to assign them IO.

We can to add a parameter "disable-io" to PCIe ports to disable IO support.
It will work by making IO base/limit registers read-only so both firmware
and guest OSes will comply.

--- Additional comment from RHEL Product and Program Management on 2016-06-09 08:04:38 EDT ---

Since this bug report was entered in bugzilla, the release flag has been
set to ? to ensure that it is properly evaluated for this release.

--- Additional comment from Marcel Apfelbaum on 2016-06-23 06:07:07 EDT ---

We intend to provide a 'thin' version of Q35 for 7.3 to be used mainly with
virtio devices which are PCIe, the IO limitation will not be an issue.


Add libvirt support for the device command line parameters.

Comment 1 Marcel Apfelbaum 2017-01-05 19:30:27 UTC
By default libvirt will disable IO space support for PCI Express Root Ports since all PCI Express devices should work just fine using only MMIO.

However there are cases when we do want to enable IO space support:
1. We might want to assign a phys device (using vfio) and the device is not spec
   compliant and requires IO ports to function properly.
2. We might have a guest firmware with an ancient driver knowing only IO ports.

Decisions like this cannot be done by libvirt, we need the upper layer to decide which PCI Express Root Ports need to have IO ports enabled.

Comment 2 Stephen Gordon 2017-01-26 21:20:55 UTC
Marcel can you provide some detail on what the actual use case is driving this?

Comment 3 Marcel Apfelbaum 2017-02-13 13:46:09 UTC
(In reply to Stephen Gordon from comment #2)
> Marcel can you provide some detail on what the actual use case is driving
> this?

Hi Stephen,
I am sorry for the delayed response, I was in PTO.

The problem is the IO ports space is rather limited and for PCI Express machines each PCI Express Root Port allows only to a single (multi-function) device. Since each such port is actually a PCI bridge, the guest firmware and after that the guest OS will try to allocate some IO range for each.

That means that after ~10 PCI root ports are used all IO ports space is exhausted and then:
 1. The firmware may panic and halt (SeaBIOS does that).
 2. The OS may try to assign the IO space by itself and in the best scenario we only see some warnings on logs.

Since all PCI Express devices should be able to work without using IO ports we can instruct the PCI Express Root Ports to disable IO ports by default.

The interesting part is what happens if we need to plug an assigned device (vfio) to a Root Port and it requires IO ports to function properly?

Since now the Root Ports have IO space disabled by default, we need to 'enable' it by setting the corresponding property.

Thanks,
Marcel

Comment 4 Stephen Finucane 2018-08-28 09:49:19 UTC
There's no clear explanation as to what use cases this feature will resolve. Until such a time as this is provided, I'm going to mark this as closed.


Note You need to log in before you can comment on or make changes to this bug.