[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Qemu-ppc] [PATCH v4 0/1] Exploit settable KVM_CAP_PPC_SMT

From: Sam Bobroff
Subject: [Qemu-ppc] [PATCH v4 0/1] Exploit settable KVM_CAP_PPC_SMT
Date: Fri, 18 Aug 2017 15:50:16 +1000
User-agent: NeoMutt/20170113 (1.7.2)

Hello QEMU PPC people,

This is v3, it is only a single patch now.

My core objective with this patch is to provide a way for QEMU to configure the
newly writeable KVM capability 'KVM_CAP_PPC_SMT', because without it Power 9
hosts can only run VMs with a single thread per core. (With this capability they
are able to run VMs with 1, 2, 4 or 8 threads per core.) KVM also now contains a
new read-only property ('KVM_CAP_PPC_SMT_POSSIBLE') to expose the possible
valid values of KVM_CAP_PPC_SMT, although this is only used in a hint to the
user at this stage. This new capability is already upstream and the QEMU
headers have already been updated to include it.

The new way KVM_CAP_PPC_SMT works is that, when set, it causes KVM to act as if
the host's native number of threads per core were the value of the capability.

I've implemented this by adding a new property to pseries machines, which is
stored in sPAPRMachineState as 'vsmt'. This provides a way to set it from the
command line (and a way to add it to the VMState if we later decide to do so)
and makes the property available only when using SPAPR (pseries) machines. I
use this value to call in to KVM and set the capability when necessary.

For pseries machines, the vsmt value will be a duplicate of the KVM capability
value, and in version 1 I tried to remove this duplication by replacing
cap_ppc_smt completely with references to spapr->vsmt. Unfortunately, that
forced generic code (e.g. translate_init.c) to have knowledge of the
sPAPRMachine state which didn't seem conceptually clean. In version 2, I've
left the capability on the KVM side which keeps the KVM and generic code clear
of SPAPR concepts. I think on the whole this is a better solution.

* I've moved the code that validates smp_threads out of ppc_cpu_realizefn()
  because it only needs to be done once, not once per CPU.

Patch set changelog follows:

====== Version 3 -> version 4: ======

Patch 1/1: PPC: KVM: Support machine option to set VSMT mode
* Use "pseries" as the machine name in some hint messages rather than the
  specific instance name, as that's sufficient.
* Merged show_vsmt_possible() and kvmppc_smt_possible() into a single function
  and moved it into kvm.c.
* Altered spapr_set_vsmt_mode() to take an error context. Note that this had
  some follow-on effects:
    * A local error context had to be added because I'm adding hints to the
    * show_vsmt_possible() now appends an error hint.

Sam Bobroff (1):
  PPC: KVM: Support machine option to set VSMT mode

 hw/ppc/spapr.c              | 75 +++++++++++++++++++++++++++++++++++++++++++++
 include/hw/ppc/spapr.h      |  1 +
 target/ppc/kvm.c            | 39 ++++++++++++++++++++++-
 target/ppc/kvm_ppc.h        | 12 ++++++++
 target/ppc/translate_init.c | 14 ---------
 5 files changed, 126 insertions(+), 15 deletions(-)


reply via email to

[Prev in Thread] Current Thread [Next in Thread]