Re: [Qemu-devel] [PATCH v4] ppc/spapr: Implement H_RANDOM hypercall in QEMU

[David Gibson]

On Mon, Sep 21, 2015 at 10:37:28AM +0200, Greg Kurz wrote:
> On Mon, 21 Sep 2015 10:26:52 +0200
> Thomas Huth  wrote:
> 
> > On 21/09/15 10:01, Greg Kurz wrote:
> > > On Mon, 21 Sep 2015 12:10:00 +1000
> > > David Gibson  wrote:
> > > 
> > >> On Fri, Sep 18, 2015 at 11:05:52AM +0200, Greg Kurz wrote:
> > >>> On Thu, 17 Sep 2015 10:49:41 +0200
> > >>> Thomas Huth  wrote:
> > >>>
> >  The PAPR interface defines a hypercall to pass high-quality
> >  hardware generated random numbers to guests. Recent kernels can
> >  already provide this hypercall to the guest if the right hardware
> >  random number generator is available. But in case the user wants
> >  to use another source like EGD, or QEMU is running with an older
> >  kernel, we should also have this call in QEMU, so that guests that
> >  do not support virtio-rng yet can get good random numbers, too.
> > 
> >  This patch now adds a new pseudo-device to QEMU that either
> >  directly provides this hypercall to the guest or is able to
> >  enable the in-kernel hypercall if available.
> > ...
> > >>> It is a good thing that the user can choose between in-kernel and 
> > >>> backend,
> > >>> and this patch does the work.
> > >>>
> > >>> This being said, I am not sure about the use case where a user has a 
> > >>> hwrng
> > >>> capable platform and wants to run guests without any hwrng support at 
> > >>> all is
> > >>> an appropriate default behavior... I guess we will find more users that 
> > >>> want
> > >>> in-kernel being the default if it is available.
> > >>>
> > >>> The patch below modifies yours to do just this: the pseudo-device is 
> > >>> only
> > >>> created if hwrng is present and not already created.
> > >>
> > >> I have mixed feelings about this.  On the one hand, I agree that it
> > >> would be nice to allow H_RANDOM support by default.  On the other hand
> > >> the patch below leaves no way to turn it off for testing purposes.  It
> > >> also adds another place where the guest hardware depends on the host
> > >> configuration, which adds to the already substantial mess of ensuring
> > >> that source and destination hardware configuration matches for
> > >> migration.
> > > 
> > > Yeah, describing the guest hw is really essential for migration... this
> > > is best addressed at the libvirt level with a full XML description of
> > > the machine... but FWIW if we are talking about running pseries on a
> > > POWER8 or newer host, I am not aware about "hwrng-less" boards... but
> > > I am probably missing something :)
> > 
> > Maybe it would be at least ok to enable it by default as long as
> > "-nodefaults" has not been specified as command line option?

I like that in principle, but the -nodefaults option isn't exposed
outside vl.c

> It makes a lot of sense indeed. I guess David should take your patch
> as it is now and the default behavior could be a follow up.

That's the plan.  I've already  taken the base patch.

-- 
David Gibson| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
| _way_ _around_!
http://www.ozlabs.org/~dgibson


pgpfA9hFQZ4X9.pgp
Description: PGP signature

Re: [Qemu-devel] [PATCH v4] ppc/spapr: Implement H_RANDOM hypercall in QEMU

[Eric Blake]

On 09/21/2015 12:00 AM, Thomas Huth wrote:

>>> This being said, I am not sure about the use case where a user has a hwrng
>>> capable platform and wants to run guests without any hwrng support at all is
>>> an appropriate default behavior... I guess we will find more users that want
>>> in-kernel being the default if it is available.
>>>
>>> The patch below modifies yours to do just this: the pseudo-device is only
>>> created if hwrng is present and not already created.
>>
>> I have mixed feelings about this.  On the one hand, I agree that it
>> would be nice to allow H_RANDOM support by default.  On the other hand
>> the patch below leaves no way to turn it off for testing purposes.  It
>> also adds another place where the guest hardware depends on the host
>> configuration, which adds to the already substantial mess of ensuring
>> that source and destination hardware configuration matches for
>> migration.
> 
> I thought about this question on the weekend and came to the same
> conclusion. I think if we want to enable this by default, it likely
> should rather be done at the libvirt level instead?
> 

Adding hardware by default to existing machine types has been a bane to
libvirt usage in the past.  If upgrading from an old qemu to a new one
suddenly turns on new guest-visible hardware with no change to the
command line, then libvirt has a much harder time migrating that guest.
 It's okay to have a new machine type turn on a feature by default, and
to have knobs so that the feature can be turned on even for older
machine types, but experience has shown that any new feature MUST come
with knobs and a way to learn if the feature can be turned on/off,
rather than just blindly assuming that turning it on is the right thing.

-- 
Eric Blake   eblake redhat com+1-919-301-3266
Libvirt virtualization library http://libvirt.org



signature.asc
Description: OpenPGP digital signature

Re: [Qemu-devel] [PATCH v4] ppc/spapr: Implement H_RANDOM hypercall in QEMU

[Thomas Huth]

On 21/09/15 10:01, Greg Kurz wrote:
> On Mon, 21 Sep 2015 12:10:00 +1000
> David Gibson  wrote:
> 
>> On Fri, Sep 18, 2015 at 11:05:52AM +0200, Greg Kurz wrote:
>>> On Thu, 17 Sep 2015 10:49:41 +0200
>>> Thomas Huth  wrote:
>>>
 The PAPR interface defines a hypercall to pass high-quality
 hardware generated random numbers to guests. Recent kernels can
 already provide this hypercall to the guest if the right hardware
 random number generator is available. But in case the user wants
 to use another source like EGD, or QEMU is running with an older
 kernel, we should also have this call in QEMU, so that guests that
 do not support virtio-rng yet can get good random numbers, too.

 This patch now adds a new pseudo-device to QEMU that either
 directly provides this hypercall to the guest or is able to
 enable the in-kernel hypercall if available.
...
>>> It is a good thing that the user can choose between in-kernel and backend,
>>> and this patch does the work.
>>>
>>> This being said, I am not sure about the use case where a user has a hwrng
>>> capable platform and wants to run guests without any hwrng support at all is
>>> an appropriate default behavior... I guess we will find more users that want
>>> in-kernel being the default if it is available.
>>>
>>> The patch below modifies yours to do just this: the pseudo-device is only
>>> created if hwrng is present and not already created.
>>
>> I have mixed feelings about this.  On the one hand, I agree that it
>> would be nice to allow H_RANDOM support by default.  On the other hand
>> the patch below leaves no way to turn it off for testing purposes.  It
>> also adds another place where the guest hardware depends on the host
>> configuration, which adds to the already substantial mess of ensuring
>> that source and destination hardware configuration matches for
>> migration.
> 
> Yeah, describing the guest hw is really essential for migration... this
> is best addressed at the libvirt level with a full XML description of
> the machine... but FWIW if we are talking about running pseries on a
> POWER8 or newer host, I am not aware about "hwrng-less" boards... but
> I am probably missing something :)

Maybe it would be at least ok to enable it by default as long as
"-nodefaults" has not been specified as command line option?

> Back to Thomas' patch, it does the job and brings H_RANDOM, which is
> currently missing.
> 
> Acked-by: Greg Kurz 
> 
> I could test both use-kvm and backend flavors (including migration).
> 
> Tested-by: Greg Kurz 

Thanks!

 Thomas




signature.asc
Description: OpenPGP digital signature