https://bugzilla.kernel.org/show_bug.cgi?id=209253 --- Comment #11 from Alex Williamson (alex.williamson@xxxxxxxxxx) --- I have reproduced, I don't know the solution yet, but I think I have a workaround, at least for occurrences like shown in comment 1. These seem to be generated by the GeForce MSI quirk, which uses an ioeventfd for the re-arming of the MSI. Perhaps we haven't unwound out of the previous MMIO write before another is triggered, but we should be running with IRQs disabled... I'm not sure. We can disable this ioeventfd in QEMU though at a slight interrupt latency cost while further debugging. libvirt xml changes: <domain type='kvm' xmlns:qemu='http://libvirt.org/schemas/domain/qemu/1.0'> ... <qemu:commandline> <qemu:arg value='-global'/> <qemu:arg value='vfio-pci.x-no-vfio-ioeventfd=on'/> </qemu:commandline> </domain> IOW, replace the first line to include the schema, otherwise libvirt will ignore and remove the addition, then add the <qemu:commandline> stanza at the end of the <domain> definition. If using QEMU command line directly you can similarly just add "-global vfio-pci.x-no-vfio-ioeventfd=on". -- You are receiving this mail because: You are watching the assignee of the bug.
