Re: [PATCH RFC 8/9] KVM: Add dirty-ring-size property

[Dr. David Alan Gilbert]

* Peter Xu (address@hidden) wrote:
> On Wed, Mar 25, 2020 at 08:00:31PM +0000, Dr. David Alan Gilbert wrote:
> > > @@ -2077,6 +2079,33 @@ static int kvm_init(MachineState *ms)
> > >      s->memory_listener.listener.coalesced_io_add = 
> > > kvm_coalesce_mmio_region;
> > >      s->memory_listener.listener.coalesced_io_del = 
> > > kvm_uncoalesce_mmio_region;
> > >  
> > > +    /*
> > > +     * Enable KVM dirty ring if supported, otherwise fall back to
> > > +     * dirty logging mode
> > > +     */
> > > +    if (s->kvm_dirty_ring_size > 0) {
> > > +        /* Read the max supported pages */
> > > +        ret = kvm_vm_check_extension(kvm_state, KVM_CAP_DIRTY_LOG_RING);
> > > +        if (ret > 0) {
> > > +            if (s->kvm_dirty_ring_size > ret) {
> > > +                error_report("KVM dirty ring size %d too big (maximum is 
> > > %d). "
> > > +                             "Please use a smaller value.",
> > > +                             s->kvm_dirty_ring_size, ret);
> > > +                goto err;
> > > +            }
> > > +
> > > +            ret = kvm_vm_enable_cap(s, KVM_CAP_DIRTY_LOG_RING, 0,
> > > +                                    s->kvm_dirty_ring_size);
> > > +            if (ret) {
> > > +                error_report("Enabling of KVM dirty ring failed: %d", 
> > > ret);
> > > +                goto err;
> > > +            }
> > > +
> > > +            s->kvm_dirty_gfn_count =
> > > +                s->kvm_dirty_ring_size / sizeof(struct kvm_dirty_gfn);
> > 
> > What happens if I was to pass dirty-ring-size=1 ?
> > Then the count would be 0 and things would get upset somewhere?
> > Do you need to check for a minimum postive value?
> 
> The above kvm_vm_enable_cap() should fail directly and QEMU will stop.
> Yes we should check it, but kernel will do that in all cases, so I
> just didn't do that explicitly again in the userspace.

We probably should have that check since you can give them a more
obvious error message.

> I was planning this to be an advanced feature so the user of this
> parameter should know the rules to pass values in.

Advanced users just make advanced mistakes :-)

I did wonder if perhaps this option should be a count of entries rather
than a byte size.

> Of course we can introduce another global knob to enable/disable this
> feature as a whole, then I can offer a default value for the size
> parameter.  I just didn't bother that in this RFC version, but I can
> switch to that if that's preferred.
> 
> [...]
> 
> > > @@ -3065,6 +3123,12 @@ static void kvm_accel_class_init(ObjectClass *oc, 
> > > void *data)
> > >          NULL, NULL, &error_abort);
> > >      object_class_property_set_description(oc, "kvm-shadow-mem",
> > >          "KVM shadow MMU size", &error_abort);
> > > +
> > > +    object_class_property_add(oc, "dirty-ring-size", "int",
> > > +        kvm_get_dirty_ring_size, kvm_set_dirty_ring_size,
> > > +        NULL, NULL, &error_abort);
> > 
> > I don't think someone passing in a non-number should cause an abort;
> > it should exit, but I don't think it should abort/core.
> 
> It won't - &error_abort is for the operation to add the property.  It
> should never fail.
> 
> User illegal input will look like this (a graceful exit):
> 
> $ ./x86_64-softmmu/qemu-system-x86_64 -accel kvm,dirty-ring-size=a
> qemu-system-x86_64: -accel kvm,dirty-ring-size=a: Parameter 'dirty-ring-size' 
> expects int64

OK good.

> > 
> > > +    object_class_property_set_description(oc, "dirty-ring-size",
> > > +        "KVM dirty ring size (<=0 to disable)", &error_abort);
> > >  }
> > >  
> > >  static const TypeInfo kvm_accel_type = {
> > > diff --git a/qemu-options.hx b/qemu-options.hx
> > > index 224a8e8712..140bd38726 100644
> > > --- a/qemu-options.hx
> > > +++ b/qemu-options.hx
> > > @@ -119,6 +119,7 @@ DEF("accel", HAS_ARG, QEMU_OPTION_accel,
> > >      "                kernel-irqchip=on|off|split controls accelerated 
> > > irqchip support (default=on)\n"
> > >      "                kvm-shadow-mem=size of KVM shadow MMU in bytes\n"
> > >      "                tb-size=n (TCG translation block cache size)\n"
> > > +    "                dirty-ring-size=n (KVM dirty ring size in Bytes)\n"
> > >      "                thread=single|multi (enable multi-threaded TCG)\n", 
> > > QEMU_ARCH_ALL)
> > >  STEXI
> > >  @item -accel @var{name}[,prop=@var{value}[,...]]
> > > @@ -140,6 +141,8 @@ irqchip completely is not recommended except for 
> > > debugging purposes.
> > >  Defines the size of the KVM shadow MMU.
> > >  @item tb-size=@var{n}
> > >  Controls the size (in MiB) of the TCG translation block cache.
> > > +@item dirty-ring-size=@val{n}
> > > +Controls the size (in Bytes) of KVM dirty ring (<=0 to disable).
> > 
> > I don't see the point in allowing < 0 ; I'd ban it.
> 
> Sure; I can switch to an uint64.

Great.

Dave

> Thanks,
> 
> -- 
> Peter Xu
> 
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK