[Date Prev][Date Next] [Thread Prev][Thread Next]
Re: [libvirt] [PATCH] rng: tighten up domain <controller> schema
- From: Osier Yang <jyang redhat com>
- To: libvir-list redhat com
- Subject: Re: [libvirt] [PATCH] rng: tighten up domain <controller> schema
- Date: Fri, 19 Apr 2013 16:32:54 +0800
On 18/04/13 19:59, Laine Stump wrote:
On 04/18/2013 07:27 AM, Osier Yang wrote:
On 18/04/13 19:16, Laine Stump wrote:
On 04/18/2013 05:41 AM, Martin Kletzander wrote:
On 04/18/2013 11:05 AM, Osier Yang wrote:
On 18/04/13 17:00, Martin Kletzander wrote:
On 04/18/2013 10:54 AM, Osier Yang wrote:
On 18/04/13 16:42, Martin Kletzander wrote:
On 04/18/2013 06:36 AM, Laine Stump wrote:
The rng schema for <controller> had been non-specific about which
types of controllers allowed which models, and also allowed the
num_queues attribute (since that hasn't been released yet,
rename it to "numQueues"?)
Since there's still time (the commit with that is
think we should rename it ASAP since we are using camelCase for
Apart from that, the RNG with this patch is precise according to
documentation, so ACK. I'll try to send the numQueues patch to see
I guess you mean multiple queues support for virtio network?
Regardless of which style we will use finally, FYI, "num_queues" is
used for disk.. Personally I'm fine with either, because we already
use both across.
Yes, I meant the virtio-scsi num_queues. As we're trying not to use
underscores in XML, I hope we can still switch it. I haven't
other num_queues anywhere in the code. Could you point me to the
that uses that? I'm sending the previously discussed patch in the
Except the virtio-scsi num_queues, there is no other tag for multiple
queue yet, we will need a patch to support multiple queue for the
but it's not committed yet.
It's fine to convert it now, 1.0.5 is not released yet. But is it
do, we already have many tags with underscore, which can't be changed
I believe those attributes  were created by mistake, and kept only
because of backward compatibility. I'm trying to be open-minded,
though, so I'm not forcing my patch in, but seeing it just as a
proposal. Others may have different opinions and I'm willing to
that. My first feeling, though, was that we should try to keep the
policy for as many of them as possible. OTOH, I've mistaken the
underscore with a hyphen when I remembered what Daniel told me about
I had recalled DV saying something about underscores in the names a long
time ago, and I recently asked about underscore vs. camelCase, and danpb
said the same thing. (Personally I don't have a preference one way or
the other, but if we really are trying to avoid them, now is our
I'm fine with either keeping it or changing num_queues. For long
term consistence, I agreed with having a consistent naming style
In the meantime, in other device types, we've tried to keep backend
details like this pushed into a <driver> subelement when possible, to
avoid polluting the main element (e.g. see the <driver> subelement of
<interface>). Is it worth putting this numQueues attribute in a <driver>
subelement too? Or am I just playing XML God?
Not sure if you mean the upcoming numQueues for interface. But for the
existing num_queues, it's for the virtio-scsi controller, putting it
doesn't reflect the purpose.
But isn't it a backend implementation detail of the specific SCSI
controller? In <interface> and <disk>, information that is specific to a
particular backend (and isn't generally applicable to that type of
device) is in the <driver> subelement.
This is the QEMU command line for a virtio-scsi disk: ("-device
is mapped to virtio-scsi controller in libvirt XML, with num_queues set):
-device virtio-scsi-pci,id=scsi0,num_queues=8,bus=pci.0,addr=0x3 \
-drive file=/dev/HostVG/QEMUGuest1,if=none,id=drive-scsi0-0-0-0 \
And this is the QEMU command line for a virtio disk (with event_idx set):
-drive file=/var/lib/libvirt/images/f14.img,if=none,id=drive-virtio-disk0 \
This is the properties the QEMU device "scsi-disk" supports:
% ./x86_64-softmmu/qemu-system-x86_64 -device scsi-disk,?
And the properties "virtio-blk-pci" device supports:
% ./x86_64-softmmu/qemu-system-x86_64 -device virtio-blk-pci,?
And the properties "virtio-scsi-pci" device supports:
% ./x86_64-softmmu/qemu-system-x86_64 -device virtio-scsi-pci,?
We can put things like "ioeventfd", "event_idx" in the <driver>
because of the QEMU device used for disk supports it. But for a
"num_queues" is supported in a separate device "virtio-scsi-pci"
means, from libvirt p.o.v, things like "ioevent_idx" are for disk,
is for the disk controller.
Assuming that we put "num_queues" or "numQueues" in <driver>, then we need
to find out the controller for disk when building QEMU command line, and
if it's virtio-scsi model, if not, error out, otherwise tell the
function to build the
controller device string that "num_queues" is specified, and what its
value is. Or
something similar but reversely (find out the disk associated with the
controller, check if num_queues is specified). This might be not the
but it can show putting "num_queues" in <driver> is just a straight
wrong way to go...
[Date Prev][Date Next] [Thread Prev][Thread Next]