TCT Meeting/November 2013 Minutes

From Xen

Attendees

  • Jun Nakajima
  • Daniel Kiper
  • Jan Beulich
  • Andres Lagar-Cavilla
  • Olaf Hering
  • Boris Ostrovsky
  • Konrad Wilk
  • Sherry Hurwitz
  • Ian Campbell
  • Don Dugger
  • Matt Wilson

Review Action items

  • xl features:
    • Konrad: xl doesn't disable PoD with PCI passthrough. It should do.
    • Matt: PCI AER pciback writes to xenstore, attributed to a domain. xend responds by shutting down the domain.

Pictogram voting comment 15px.png Action Matt: to raise it on the list.

Current technical challenges

  • Sherry: AMD working on adding a broadcom device for Serial over LAN. The LAN is seen by dom0 but the UART part needs to stay with Xen. People seem to be having issues with doing this on the list, hiding individual PCI device functions from dom0. Might be some more work in the generic area?
    • Konrad: Was Andrew Cooper looking at something like this?
    • Jan: Xen is not hiding devices from dom0, it is preventing dom0 from making changes to devices which it is using. So device enumeration should work, but writing to the config space should be suppressed. How device specific would it be, and how far would it go?
    • Sherry: AMD will look into this and make proposals on the list.

Coordination of future work

  • Daniel: EFI boot protocol. Is going to continue with what was discussed on the list.
  • Internships (Konrad)
    • Konrad: Xen AB has provide funding for GNOME OPW. Two candidates, one slot. Money set aside for an intern during the summer. Interns on Xen or Linux. Any project ideas (3 months) for the summer one could be posted to xen-devel.
  • SR-IOV, PCI passthrough (Konrad)
    • Konrad: PCI passthrough doesn't feel complete. Things which don't work that well. Would like to change this. Who is using this, are there known bugs. Should we slip the release?
    • Ian: Bugs or regressions?
    • Konrad: Thinks bugs
    • Sherry: What devices
    • Konrad: Network, infiniband, storage
    • Don what are the errors you are seeing?
    • Jan: Two known things are XSA-59 and -60. -60 has patches in unstable, but not backported since not yet complete. XSA-59 trying to get something from Intel.
    • Don: needs to push on security issues. But as far as functionality is working
    • Konrad: BARs above 4G doesn't work with qemu. Reloading drivers doesn't work well.
    • Jun: Patches have been sent out to support BARs above 4G.
    • Ian: Might be an issue with qemu-upstream which was band aided in 4.3 and not yet fixed;
    • Konrad: Is using qemu-traditional
    • Jan: Things either still don't work even with 4G patches or something is missing
    • Matt: Reload issues to do with MSI perhaps?
    • Don: Should post issues on xen-devel.
    • Konrad: More and more people wanting to passthrough GPUs. Lots of odd PCI cards, which don't work so well with passthrough.
    • Matt: GPUs are a special beast. Especially primary display passthrough is problematic. Secondary works "pretty well" for most GPUs. Primary requires BIOS to run in the guest, VBAR=PBAR etc etc etc
    • Sherry: The AMD GFX team have been looking at some issues with their GPUs and there may be some patches forthcoming.
    • IanC: Patches floating around in the community to fix "stuff"
    • Matt: Tha's mostly VBAR==PBAR.
    • Konrad: Should we have a goal to have these things identified and working on 4.5
    • Jun: Graphics is "special", lots of issues with firmware, BIOS, driver etc. If we want to support GFX passthrough then perhaps a whitelist of GFX controller.
    • Ian: I think there is such a list on the wiki.
    • Jun: Intel testing SR-IOV, mostly NICs. Are often broken. Need to have standard testing.
    • Ian: osstest has a PCI passthrogh NIC test which always fails, needs someone to diagnose
    • Konrad: So who cares about this stuff? Those who are interested should talk together:
      • Konrad
      • Intel (Don)
      • AMD (Sherry). Have people from GFX team working on Xen. Kelly Zyaruk?
      • IanC thinks XenServer might be interested. but cannot speak for them.
  • PCI passthrough with other tools not xl
    • use pci assign/detach etc.
    • but virt-manager also has an option to do this. libxl+libvirt could therefore drive this.

    • Pictogram voting comment 15px.png Action Konrad: to talk to Jim Fehlig
  • Grand Table Scalability (Matt)
    • Talk on list of extending the locks available.
    • Any thoughts on other locking scalability?
    • Konrad is happy to test patches.
    • Matt: Grant table lock is eclipsing any other issues for storage workloads.
    • Will iterate on the list.