[libvirt] [PATCH] rng: tighten up domain <controller> schema

Fri Apr 19 15:25:32 UTC 2013

On 04/19/2013 04:32 AM, Osier Yang wrote:
> 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,
>>>>>>>>>> should we
>>>>>>>>>> rename it to "numQueues"?)
>>>>>>>>> Since there's still time (the commit with that is
>>>>>>>>> v1.0.4-65-gd4bf0a9), I
>>>>>>>>> think we should rename it ASAP since we are using camelCase for
>>>>>>>>> all the
>>>>>>>>> attribute names.
>>>>>>>>>
>>>>>>>>> Apart from that, the RNG with this patch is precise according to
>>>>>>>>> the
>>>>>>>>> documentation, so ACK.  I'll try to send the numQueues patch
>>>>>>>>> to see
>>>>>>>>> what
>>>>>>>>> others think.
>>>>>>>> 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
>>>>>>> found any
>>>>>>> other num_queues anywhere in the code.  Could you point me to the
>>>>>>> commit
>>>>>>> that uses that?  I'm sending the previously discussed patch in the
>>>>>>> meantime.
>>>>>>>
>>>>>> 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
>>>>>> network,
>>>>>> but it's not committed yet.
>>>>>>
>>>>>> It's fine to convert it now, 1.0.5 is not released yet. But is it
>>>>>> deserved to
>>>>>> do, we already have many tags with underscore, which can't be
>>>>>> changed
>>>>>> for back-compat.
>>>>>>
>>>>> I believe those attributes [1] 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
>>>>> discuss
>>>>> that.  My first feeling, though, was that we should try to keep the
>>>>> same
>>>>> 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
>>>>> attributes [2].
>>>> 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
>>>> chance).
>>> I'm fine with either keeping it or changing num_queues. For long
>>> term consistence, I agreed with having a consistent naming style
>>> is nice.
>>>
>>>> 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
>>> in <driver>
>>> 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
> virtio-scsi-pci"
> 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 \
> -usb \
> -drive file=/dev/HostVG/QEMUGuest1,if=none,id=drive-scsi0-0-0-0 \
> -device
> scsi-disk,bus=scsi0.0,channel=0,scsi-id=0,lun=0,drive=drive-scsi0-0-0-0,id=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 \
> -device
> virtio-blk-pci,event_idx=on,scsi=off,bus=pci.0,addr=0x4,drive=drive-virtio-disk0,id=virtio-disk0
> \
> </...>
>
> This is the properties the QEMU device "scsi-disk" supports:
>
> % ./x86_64-softmmu/qemu-system-x86_64 -device scsi-disk,?
> scsi-disk.drive=drive
> scsi-disk.logical_block_size=blocksize
> scsi-disk.physical_block_size=blocksize
> scsi-disk.min_io_size=uint16
> scsi-disk.opt_io_size=uint32
> scsi-disk.bootindex=int32
> scsi-disk.discard_granularity=uint32
> scsi-disk.ver=string
> scsi-disk.serial=string
> scsi-disk.vendor=string
> scsi-disk.product=string
> scsi-disk.removable=on/off
> scsi-disk.dpofua=on/off
> scsi-disk.wwn=hex64
> scsi-disk.channel=uint32
> scsi-disk.scsi-id=uint32
> scsi-disk.lun=uint32
>
> And the properties "virtio-blk-pci" device supports:
>
> % ./x86_64-softmmu/qemu-system-x86_64 -device virtio-blk-pci,?
> virtio-blk-pci.class=hex32
> virtio-blk-pci.ioeventfd=on/off
> virtio-blk-pci.vectors=uint32
> virtio-blk-pci.indirect_desc=on/off
> virtio-blk-pci.event_idx=on/off
> virtio-blk-pci.drive=drive
> virtio-blk-pci.logical_block_size=blocksize
> virtio-blk-pci.physical_block_size=blocksize
> virtio-blk-pci.min_io_size=uint16
> virtio-blk-pci.opt_io_size=uint32
> virtio-blk-pci.bootindex=int32
> virtio-blk-pci.discard_granularity=uint32
> virtio-blk-pci.cyls=uint32
> virtio-blk-pci.heads=uint32
> virtio-blk-pci.secs=uint32
> virtio-blk-pci.serial=string
> virtio-blk-pci.config-wce=on/off
> virtio-blk-pci.scsi=on/off
> virtio-blk-pci.addr=pci-devfn
> virtio-blk-pci.romfile=string
> virtio-blk-pci.rombar=uint32
> virtio-blk-pci.multifunction=on/off
> virtio-blk-pci.command_serr_enable=on/off
>
> And the properties "virtio-scsi-pci" device supports:
>
> % ./x86_64-softmmu/qemu-system-x86_64 -device virtio-scsi-pci,?
> virtio-scsi-pci.ioeventfd=on/off
> virtio-scsi-pci.vectors=uint32
> virtio-scsi-pci.indirect_desc=on/off
> virtio-scsi-pci.event_idx=on/off
> virtio-scsi-pci.hotplug=on/off
> virtio-scsi-pci.param_change=on/off
> virtio-scsi-pci.num_queues=uint32
> virtio-scsi-pci.max_sectors=uint32
> virtio-scsi-pci.cmd_per_lun=uint32
> virtio-scsi-pci.addr=pci-devfn
> virtio-scsi-pci.romfile=string
> virtio-scsi-pci.rombar=uint32
> virtio-scsi-pci.multifunction=on/off
> virtio-scsi-pci.command_serr_enable=on/off
>
> We can put things like "ioeventfd", "event_idx" in the <driver>
> subelement, is
> because of the QEMU device used for disk supports it. But for a
> virtio-scsi disk,
> "num_queues" is supported in a separate device "virtio-scsi-pci"
> instead.. That
> means, from libvirt p.o.v,  things like "ioevent_idx" are for disk,
> "num_queues"
> 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 check
> 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
> virtio-scsi
> controller, check if num_queues is specified). This might be not the
> exact process,
> but it can show putting "num_queues" in <driver> is just a straight
> wrong way to go...

Wait. So you're saying that num_queues is a property of the *controller*
and not of the individual disk, but you've put the config option in the
<disk> rather than the <controller>? Why would you do that? If it's a
property of the controller, put the tuning parameter in <controller>.
Otherwise, what do you do when one <disk> is configured for
num_queues=10 and another disk on the same controller is configured for
num_queues=2?

(And even if you didn't move the config to <controller> (where it seems
to me it belongs), moving to the <driver> subelement would still be
appropriate - it's still a "backend-specific tuning parameter").