Message ID | 157478677206.67101.11965554900825793694.stgit@bahia.tlslab.ibm.com |
---|---|
Headers | show |
Series | spapr: Use less XIVE HW resources in KVM | expand |
On Tue, Nov 26, 2019 at 05:46:12PM +0100, Greg Kurz wrote: > On POWER9 systems, the XICS-on-XIVE and XIVE KVM devices currently > allocate a bunch of VPs in the XIVE HW to accomodate the highest > VCPU id that may be possibly used in a VM. This limits the number > of VMs that can run with an in-kernel interrupt controller to 63 > per POWER9 chip, irrespectively of its number of HW threads, eg. > up to 96 on a POWER9 Nimbus socket. This is an unfortunate waste > of scarce HW resources since a typical VM doesn't need that much > VPs to run. > > This series exploits new attributes of the XICS-on-XIVE and XIVE > KVM devices that allow userspace to tune the numbers of VPs it > really needs. Support for this attributes has just reached linux > mainline. > > In combination with recent commit 29cb4187497d "spapr: Set VSMT to > smp_threads by default" which makes VCPU ids contiguous, it is > possible to start more than 200 VMs instead of 127 on an AC922 > system (two sockets, 144 HW threads). > > Changes in v2: > - add an nr_servers argument to the activate() handler instead of > a QOM property for XIVE and a XICSFabric method for XICS-on-XIVE. > - even if the last two patches haven't changed much, the code base > itself has, so I've deliberately dropped Cedric's R-b tags Applied to ppc-for-5.0, thanks.