2016-08-11 21:29+0800, Peter Xu: > Adding one extra property for intel-iommu device to decide whether we > should support EIM bit for IR. > > Now we are throwing high 24 bits of dest_id away directly. This will > cause interrupt issues with guests that: > > - enabled x2apic with cluster mode > - have more than 16 vcpus (so will have more than 1 cluster)
Good catch, it is a problem starting at 9 VCPUs, because that one is already "0x100" and we truncate to 8 bits ... > Let's make xapic the default one, and for the brave people who would > like to try EIM and know the side effects, we can do it by explicitly > enabling EIM using: This might bite us: users cannot easily tell when EIM is sane, so eim=on is going to be a hazard even after some KVM/QEMU has it fixed. I'd make make eim=on fail when logical x2APIC is broken. Right now, we can allow eim=on if maximal APIC ID < 8, because physical 0xffffffff gets truncated to 0xff and is still interpreted as a broadcast thanks to the quirk. Logical 0xff gets misinterpreted as a broadcast, but it would address all APICs anyway and KVM's lowest-priority works for broadcast too, so it is fine to do so. Not very wise, though. :) > -device intel-iommu,intremap=on,eim=on > > Even after we have x2apic support, it'll still be good if we can provide > a way to switch xapic/x2apic from QEMU side for e.g. debugging purpose, > which is an alternative for tuning guest kernel boot parameters. > > We can switch the default to "on" after x2apic fully supported. > > Signed-off-by: Peter Xu <pet...@redhat.com> > --- > This will have small conflict with Radim's patches to selectively > enable EIM. I can do a rebase when needed. This patch made me realize that broadcast quirk is not enough, so I'll do v2 on top of this.
