Re: [PATCH v1 04/10] hw/arm: xenpvh: Add support for SMP guests

[Stefano Stabellini]

On Fri, 16 Aug 2024, Edgar E. Iglesias wrote:
> On Thu, Aug 15, 2024 at 2:30 AM Stefano Stabellini <sstabellini@kernel.org> 
> wrote:
>       On Wed, 14 Aug 2024, Edgar E. Iglesias wrote:
>       > On Tue, Aug 13, 2024 at 03:52:32PM -0700, Stefano Stabellini wrote:
>       > > On Tue, 13 Aug 2024, Edgar E. Iglesias wrote:
>       > > > On Mon, Aug 12, 2024 at 06:47:17PM -0700, Stefano Stabellini 
> wrote:
>       > > > > On Mon, 12 Aug 2024, Edgar E. Iglesias wrote:
>       > > > > > From: "Edgar E. Iglesias" <edgar.iglesias@amd.com>
>       > > > > >
>       > > > > > Add SMP support for Xen PVH ARM guests. Create max_cpus ioreq
>       > > > > > servers to handle hotplug.
>       > > > > >
>       > > > > > Signed-off-by: Edgar E. Iglesias <edgar.iglesias@amd.com>
>       > > > > > ---
>       > > > > >  hw/arm/xen_arm.c | 5 +++--
>       > > > > >  1 file changed, 3 insertions(+), 2 deletions(-)
>       > > > > >
>       > > > > > diff --git a/hw/arm/xen_arm.c b/hw/arm/xen_arm.c
>       > > > > > index 5f75cc3779..ef8315969c 100644
>       > > > > > --- a/hw/arm/xen_arm.c
>       > > > > > +++ b/hw/arm/xen_arm.c
>       > > > > > @@ -173,7 +173,7 @@ static void xen_arm_init(MachineState 
> *machine)
>       > > > > > 
>       > > > > >      xen_init_ram(machine);
>       > > > > > 
>       > > > > > -    xen_register_ioreq(xam->state, machine->smp.cpus, 
> &xen_memory_listener);
>       > > > > > +    xen_register_ioreq(xam->state, machine->smp.max_cpus, 
> &xen_memory_listener);
>       > > > > > 
>       > > > > >      xen_create_virtio_mmio_devices(xam);
>       > > > > > 
>       > > > > > @@ -218,7 +218,8 @@ static void 
> xen_arm_machine_class_init(ObjectClass *oc, void *data)
>       > > > > >      MachineClass *mc = MACHINE_CLASS(oc);
>       > > > > >      mc->desc = "Xen PVH ARM machine";
>       > > > > >      mc->init = xen_arm_init;
>       > > > > > -    mc->max_cpus = 1;
>       > > > > > +    /* MAX number of vcpus supported by Xen.  */
>       > > > > > +    mc->max_cpus = GUEST_MAX_VCPUS;
>       > > > >
>       > > > > Will this cause allocations of data structures with 128 
> elements?
>       > > > > Looking at hw/xen/xen-hvm-common.c:xen_do_ioreq_register it 
> seems
>       > > > > possible? Or hw/xen/xen-hvm-common.c:xen_do_ioreq_register is 
> called
>       > > >
>       > > > Yes, in theory there's probably overhead with this but as you 
> correctly
>       > > > noted below, a PVH aware xl will set the max_cpus option to a 
> lower value.
>       > > >
>       > > > With a non-pvh aware xl, I was a little worried about the overhead
>       > > > but I couldn't see any visible slow-down on ARM neither in boot 
> or in network
>       > > > performance (I didn't run very sophisticated benchmarks).
>       > > 
>       > > What do you mean by "non-pvh aware xl"? All useful versions of xl
>       > > support pvh?
>       >
>       >
>       > I mean an xl without our PVH patches merged.
>       > xl in upstream doesn't know much about PVH yet.
>       > Even for ARM, we're still carrying significant patches in our tree.
> 
>       Oh I see. In that case, I don't think we need to support "non-pvh aware 
> xl".
> 
> 
>       > > > > later on with the precise vCPU value which should be provided 
> to QEMU
>       > > > > via the -smp command line option
>       > > > > 
> (tools/libs/light/libxl_dm.c:libxl__build_device_model_args_new)?
>       > > >
>       > > > Yes, a pvh aware xl will for example pass -smp 2,maxcpus=4 based 
> on
>       > > > values from the xl.cfg. If the user doesn't set maxvcpus in 
> xl.cfg, xl
>       > > > will set maxvcpus to the same value as vcpus.
>       > >
>       > > OK good. In that case if this is just an initial value meant to be
>       > > overwritten, I think it is best to keep it as 1.
>       >
>       > Sorry but that won't work. I think the confusion here may be that
>       > it's easy to mix up mc->max_cpus and machine->smp.max_cpus, these are
>       > not the same. They have different purposes.
>       >
>       > I'll try to clarify the 3 values in play.
>       >
>       > machine-smp.cpus:
>       > Number of guest vcpus active at boot.
>       > Passed to QEMU via the -smp command-line option.
>       > We don't use this value in QEMU's ARM PVH machines.
>       >
>       > machine->smp.max_cpus:
>       > Max number of vcpus that the guest can use (equal or larger than 
> machine-smp.cpus).
>       > Will be set by xl via the "-smp X,maxcpus=Y" command-line option to 
> QEMU.
>       > Taken from maxvcpus from xl.cfg, same as XEN_DMOP_nr_vcpus.
>       > This is what we use for xen_register_ioreq().
>       >
>       > mc->max_cpus:
>       > Absolute MAX in QEMU used to cap the -smp command-line options.
>       > If xl tries to set -smp (machine->smp.max_cpus) larger than this, 
> QEMU will bail out.
>       > Used to setup xen_register_ioreq() ONLY if -smp maxcpus was NOT set 
> (i.e by a non PVH aware xl).
>       > Cannot be 1 because that would limit QEMU to MAX 1 vcpu.
>       >
>       > I guess we could set mc->max_cpus to what XEN_DMOP_nr_vcpus returns 
> but I'll
>       > have to check if we can even issue that hypercall this early in QEMU 
> since
>       > mc->max_cpus is setup before we even parse the machine options. We may
>       > not yet know what domid we're attaching to yet.
> 
>       If mc->max_cpus is the absolute max and it will not be used if -smp is
>       passed to QEMU, then I think it is OK to use GUEST_MAX_VCPUS
> 
> Looking at this a little more. If users (xl) don't pass an -smp option we 
> actually default to smp.max_cpus=1.
> So, another option is to simply remove the upper limit in QEMU (e.g we can 
> set mc->max_cpus to something very large like UINT32_MAX).
> That would avoid early hypercalls, avoid using GUEST_MAX_VCPUS and always let 
> xl dictate the max_cpus value using the -smp cmdline option. 

As the expectation is that there will be always a smp.max_cpus option
passed to QEMU, I would avoid an extra early hypercall.

For the initial value, I would use something static and large, but not
unreasonably large as UINT32_MAX to be more resilient in (erroneous)
cases where smp.max_cpus is not passed.

So I would initialize it to GUEST_MAX_VCPUS, or if we don't want to use
GUEST_MAX_VCPUS, something equivalent in the 64-256 range.

Alternative we can have a runtime check and exit with a warning if
smp.max_cpus is not set.