Re: [Qemu-devel] [PATCH v6 09/18] hw/arm/virt: Implement kvm_type function for 4.0 machine

[Igor Mammedov <imammedo@redhat.com>]

On Mon, 18 Feb 2019 22:29:40 +0100
Auger Eric <eric.auger@redhat.com> wrote:

> Hi Peter,
> 
> On 2/14/19 6:29 PM, Peter Maydell wrote:
> > On Tue, 5 Feb 2019 at 17:33, Eric Auger <eric.auger@redhat.com> wrote:  
> >>
> >> This patch implements the machine class kvm_type() callback.
> >> It returns the max IPA shift needed to implement the whole GPA
> >> range including the RAM and IO regions located beyond.
> >> The returned value in passed though the KVM_CREATE_VM ioctl and
> >> this allows KVM to set the stage2 tables dynamically.
> >>
> >> At this stage the RAM limit still is limited to 255GB.
> >>
> >> Setting all the existing highmem IO regions beyond the RAM
> >> allows to have a single contiguous RAM region (initial RAM and
> >> possible hotpluggable device memory). That way we do not need
> >> to do invasive changes in the EDK2 FW to support a dynamic
> >> RAM base.
> >>
> >> Signed-off-by: Eric Auger <eric.auger@redhat.com>
> >>
> >> ---
> >>
> >> v5 -> v6:
> >> - add some comments
> >> - high IO region cannot start before 256GiB
> >> ---
> >>  hw/arm/virt.c         | 52 +++++++++++++++++++++++++++++++++++++++++--
> >>  include/hw/arm/virt.h |  2 ++
> >>  2 files changed, 52 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/hw/arm/virt.c b/hw/arm/virt.c
> >> index 2b15839d0b..b90ffc2e5d 100644
> >> --- a/hw/arm/virt.c
> >> +++ b/hw/arm/virt.c
> >> @@ -1366,6 +1366,7 @@ static uint64_t virt_cpu_mp_affinity(VirtMachineState *vms, int idx)
> >>
> >>  static void virt_set_memmap(VirtMachineState *vms)
> >>  {
> >> +    MachineState *ms = MACHINE(vms);
> >>      hwaddr base;
> >>      int i;
> >>
> >> @@ -1375,7 +1376,17 @@ static void virt_set_memmap(VirtMachineState *vms)
> >>          vms->memmap[i] = a15memmap[i];
> >>      }
> >>
> >> -    vms->high_io_base = 256 * GiB; /* Top of the legacy initial RAM region */
> >> +    /*
> >> +     * We now compute the base of the high IO region depending on the
> >> +     * amount of initial and device memory. The device memory start/size
> >> +     * is aligned on 1GiB. We never put the high IO region below 256GiB
> >> +     * so that if maxram_size is < 255GiB we keep the legacy memory map
> >> +     */
> >> +    vms->high_io_base = ROUND_UP(GiB + ms->ram_size, GiB) +
> >> +                        ROUND_UP(ms->maxram_size - ms->ram_size, GiB);  
> > 
> > I don't understand this expression...  
> My intent was to align the start of the device memory on a GiB boundary,
> just after the initial RAM (ram_size). And then align the floating IO
> region on a GiB boundary after the device memory (of size
> ms->maxram_size - ms->ram_size). What do I miss?

It's not obvious what "GiB +  ms->ram_size" means and where it comes from,
maybe substitute GiB with properly named constant/macro that's also re-used in
memmap definition so it would be obvious that's it's where initial RAM
is mapped. Also I'd move both ROUND_UPs into separate expressions using
reasonable named local vars and possible overflow checks on top of that,
so one won't have to guess that it's initial RAM end + device RAM end.

> >   
> >> +    if (vms->high_io_base < 256 * GiB) {
> >> +        vms->high_io_base = 256 * GiB;
> >> +    }
> >>      base = vms->high_io_base;
> >>
> >>      for (i = VIRT_LOWMEMMAP_LAST; i < ARRAY_SIZE(extended_memmap); i++) {
> >> @@ -1386,6 +1397,7 @@ static void virt_set_memmap(VirtMachineState *vms)
> >>          vms->memmap[i].size = size;
> >>          base += size;
> >>      }
> >> +    vms->highest_gpa = base - 1;
> >>  }
> >>
> >>  static void machvirt_init(MachineState *machine)
> >> @@ -1402,7 +1414,9 @@ static void machvirt_init(MachineState *machine)
> >>      bool firmware_loaded = bios_name || drive_get(IF_PFLASH, 0, 0);
> >>      bool aarch64 = true;
> >>
> >> -    virt_set_memmap(vms);
> >> +    if (!vms->extended_memmap) {
> >> +        virt_set_memmap(vms);
> >> +    }
> >>
> >>      /* We can probe only here because during property set
> >>       * KVM is not available yet
> >> @@ -1784,6 +1798,36 @@ static HotplugHandler *virt_machine_get_hotplug_handler(MachineState *machine,
> >>      return NULL;
> >>  }
> >>
> >> +/*
> >> + * for arm64 kvm_type [7-0] encodes the IPA size shift
> >> + */
> >> +static int virt_kvm_type(MachineState *ms, const char *type_str)
> >> +{
> >> +    VirtMachineState *vms = VIRT_MACHINE(ms);
> >> +    int max_vm_phys_shift = kvm_arm_get_max_vm_phys_shift(ms);
> >> +    int max_pa_shift;
> >> +
> >> +    vms->extended_memmap = true;
> >> +
> >> +    virt_set_memmap(vms);
> >> +
> >> +    max_pa_shift = 64 - clz64(vms->highest_gpa);
> >> +
> >> +    if (max_pa_shift > max_vm_phys_shift) {
> >> +        error_report("-m and ,maxmem option values "
> >> +                     "require an IPA range (%d bits) larger than "
> >> +                     "the one supported by the host (%d bits)",
> >> +                     max_pa_shift, max_vm_phys_shift);
> >> +       exit(1);
> >> +    }  
> > 
> > Presumably we should have some equivalent check for TCG, so
> > that we don't let the user create a setup which wants more
> > bits of physical address than the TCG CPU allows ?  
> kvm_type() sets the new memory map. For TCG we should stick to the 1TB
> GPA address space which should be consistent with the existing
> ID_AA64MMFR0_EL1 settings (arm/internals.h implements arm_pamax(ARMCPU
> *cpu) which decodes hardcoded cpu->id_aa64mmfr0).
> >   
> >> +    /*
> >> +     * By default we return 0 which corresponds to an implicit legacy
> >> +     * 40b IPA setting. Otherwise we return the actual requested IPA
> >> +     * logsize
> >> +     */
> >> +    return max_pa_shift > 40 ? max_pa_shift : 0;
> >> +}
> >> +
> >>  static void virt_machine_class_init(ObjectClass *oc, void *data)
> >>  {
> >>      MachineClass *mc = MACHINE_CLASS(oc);
> >> @@ -1808,6 +1852,7 @@ static void virt_machine_class_init(ObjectClass *oc, void *data)
> >>      mc->cpu_index_to_instance_props = virt_cpu_index_to_props;
> >>      mc->default_cpu_type = ARM_CPU_TYPE_NAME("cortex-a15");
> >>      mc->get_default_cpu_node_id = virt_get_default_cpu_node_id;
> >> +    mc->kvm_type = virt_kvm_type;
> >>      assert(!mc->get_hotplug_handler);
> >>      mc->get_hotplug_handler = virt_machine_get_hotplug_handler;
> >>      hc->plug = virt_machine_device_plug_cb;
> >> @@ -1911,6 +1956,9 @@ static void virt_machine_3_1_options(MachineClass *mc)
> >>  {
> >>      virt_machine_4_0_options(mc);
> >>      compat_props_add(mc->compat_props, hw_compat_3_1, hw_compat_3_1_len);
> >> +
> >> +    /* extended memory map is enabled from 4.0 onwards */
> >> +    mc->kvm_type = NULL;  
> > 
> > When is there a difference between setting this to NULL,
> > and setting it to virt_kvm_type but having the memory
> > size be <= 256GiB ?  
> There shouldn't be any difference. When size <= 255GiB we stick to the
> 1TB PA address space.
> > 
> > If there isn't any difference, why can't we just let the
> > pre-4.0 versions behave like the new ones? No existing
> > VM setup will have > 256GB of memory, so as long as there's
> > no behaviour change for the <=256GB case we don't need to
> > take special effort to ensure that the >256GB case continues
> > to give an error message, do we ?  
> But don't we want to forbid any pre-4.0 machvirt to run with more than
> 255GiB RAM?
Why would we if it doesn't break migration?

 
> Thanks
> 
> Eric
> >   
> >>  }
> >>  DEFINE_VIRT_MACHINE(3, 1)
> >>
> >> diff --git a/include/hw/arm/virt.h b/include/hw/arm/virt.h
> >> index 3dc7a6c5d5..c88f67a492 100644
> >> --- a/include/hw/arm/virt.h
> >> +++ b/include/hw/arm/virt.h
> >> @@ -132,6 +132,8 @@ typedef struct {
> >>      uint32_t iommu_phandle;
> >>      int psci_conduit;
> >>      hwaddr high_io_base;
> >> +    hwaddr highest_gpa;
> >> +    bool extended_memmap;
> >>  } VirtMachineState;
> >>
> >>  #define VIRT_ECAM_ID(high) (high ? VIRT_HIGH_PCIE_ECAM : VIRT_PCIE_ECAM)
> >> --
> >> 2.20.1  
> > 
> > thanks
> > -- PMM
> >   
>