On Fri, 23 Aug 2019, 0:27 Laine Stump, <laine@redhat.com> wrote:
(Adding Alex Williamson to Cc so he can correct any mistakes)

On 8/22/19 4:39 PM, Ihar Hrachyshka wrote:
> On Thu, Aug 22, 2019 at 12:01 PM Laine Stump <laine@redhat.com> wrote:
>> On 8/22/19 10:56 AM, Ihar Hrachyshka wrote:
>>> On Thu, Aug 22, 2019 at 2:24 AM Daniel P. Berrangé <berrange@redhat.com> wrote:
>>>> On Wed, Aug 21, 2019 at 01:37:21PM -0700, Ihar Hrachyshka wrote:
>>>>> Hi all,
>>>>> KubeVirt uses libvirtd to manage qemu VMs represented as Kubernetes
>>>>> API resources. In this case, libvirtd is running inside an
>>>>> unprivileged pod, with some host mounts / capabilities added to the
>>>>> pod, needed by libvirtd and other services.
>>>>> One of the capabilities libvirtd requires for successful startup
>>>>> inside a pod is SYS_RESOURCE. This capability is used to adjust
>>>>> RLIMIT_MEMLOCK ulimit value depending on devices attached to the
>>>>> managed guest, both on startup and during hotplug. AFAIU the need to
>>>>> lock the memory is to avoid pages being pushed out from RAM into swap.
>> I recall successfully testing GPU assignment from an unprivileged
>> libvirtd several years ago by setting a high enough ulimit for the uid
>> used to run libvirtd in advance (. I think we check if the current
>> setting is high enough, and don't try to set it unless we think we need to.
> The PR I linked to in the original email does just that: it starts
> libvirtd; then, if domain is going to use VFIO, sets ulimit of
> libvirtd process to VM memory size + 1Gb (mimicking libvirt code) +
> 256Mb (to stay conservative) using prlimit() syscall; then defines the
> domain.

So you're making an educated guess, which is essentially what libvirt is
doing (based on advice from other people with better information than
us, but still a guess).

>> If I understand you correctly, you're saying that in your case it's okay
>> for the memlock limit to be lower than we try to set it to, because swap
>> is disabled anyway, is that correct?
> I'm honestly not exactly sure about the reason why we need to set the
> limit, but I assume it's because of swap. I can be totally confused on
> that part though.

What I understand from an IRC conversation with Alex just now is that
increasing RLIMIT_MEMLOCK isn't done just to prevent any of the pages
being swapped out. It's done because "all GPAs (Guest Physical
Addresses) that could potentially be DMA targets need to have fixed
mappings through the iommu, therefore all need to be allocated and
mappings fixed [...] setting rlimit allows us to perform all the
necessary pins within the user's locked memory limit".

So even if swap is disabled, it still needs to be done (either by
libvirt, or by someone else who has the necessary privileges and control
over the libvirtd process).

>>>> Libvirt shouldn't set RLIMIT_MEMLOCK by default, unless there's
>>>> something in the XML that requires it - one of
>>> You are right, sorry. We add SYS_RESOURCE only for particular domains.
>>>>    - hard limit memory value is present
>>>>    - host PCI device passthrough is requested
>>> We are using passthrough
>> (If you want to make Alex happy, use the term "VFIO device assignment"
>> rather than passthrough :-).)
> Not sure who Alex is but I'll try to make everyone happy! :)

The Alex I'm referring to is the Alex I just Cc'ed. He is the VFIO

>>> to pass SR-IOV NIC VFs into guests. We also
>>> plan to do the same for GPUs in the near future.
>>   >>> I believe we would benefit from one of the following features on
>>   >>> libvirt side (or both):
>>   >>>
>>   >>> a) expose the memory lock value calculated by libvirtd through
>>   >>> libvirt ABI so that we can use it when calling prlimit() on libvirtd
>>   >>> process;
>>   >>> b) allow to disable setrlimit() calls via libvirtd config file knob
>>   >>> or domain definition.
>> (b) sounds much more reasonable, as long as qemu doesn't complain (I
>> don't know whether or not it checks)
>> Slightly related to this - I'm currently working on patches to avoid
>> making any ioctl calls that would fail in an unprivileged libvirtd when
>> using tap/macvtap devices.

This is music to my ears, great to hear.

ATM, I'm doing this by adding an attribute
>> "unmanaged='yes'" to the interface <target> element. The idea is that if
>> someone sets unmanaged='yes', they're stating that the caller (i.e.
>> kubevirt) is responsible for all device setup, and that libvirt should
>> just use it without further setup. A similar approach could be applied
>> to hostdev devices - if unmanaged is set, we assume that the caller has
>> done everything to make the associated device usable.
>> (Of course this all makes me realize the inanity of adding a <target
>> dev='blah' unmanaged='yes'/> for interfaces when hostdevs already have
>> <hostdev managed='yes'> and <interface type='hostdev' managed='yes'>. So
>> to prevent setting the locklimit for hostdev, would we make a new
>> setting like <hostdev managed='no-never-not-even-a-tiny-bit'>? Sigh. I
>> *hate* trying to make config consistent :-/)

Sounds tough indeed. I'd try to avoid negatively-named knobs. managed=no is simpler to perceive than unmanaged=yes. It may be just me, but I'd even assume managed=no if the target dev name is specified. If libvirt manages the tap device, it should create a fresh one, too. But all of this is a big digression.

>> (alternately, we could just automatically fail the attempt to set the
>> lock limit in a graceful manner and allow the guest to continue)
> If that's something maintainers feel good about, I am all for it since
> it simplifies the implementation.

Well, after talking to Alex, I think that since a) libvirt only attempts
to increase the limit after determining that it isn't already high
enough, and b) if it isn't high enough and we can't increase it, then
qemu is going to fail anyway, that c) we can't just fail gracefully and

So *somebody* needs to increase the limit, and if you want libvirt to be
unprivileged, that means it needs to be you doing the increase. And
since the amount that libvirt increases it is just some number based on
oral folklore (and not on a specific value we learn by querying
somewhere), I don't think it's worthwhile figuring out some way for
libvirt to report it via an official API - that would end up just being

"Hey, you know that number that you guys are just making a guess about
based on some advice someone gave you once? Yeah, send me *that* number
so I can claim to be basing my actions on real science instead of
slightly educated voodoo! K THX BYE!" :-)

Well, it's more like: "you know that voodoo you do to guess the number? If you ever educate yourself about it, e.g by querying qemu, send me *that* number. I'd rather not think about it ever again, BYE."

>> BTW, I'm guessing that you use <hostdev> to assign the SRIOV VFs rather
>> than <interface type='hostdev'>, correct? The latter would require that
>> you have enough capabilities to set MAC addresses on the VFs (that's the
>> entire point of using <interface type='hostdev'> instead of plain <hostdev>)
> Yes, we use <hostdev> exactly because interface sets MAC address: in
> kubevirt scenario, the container that is running libvirtd has its own
> network namespace and doesn't have access to PF to set the VF MAC
> address on. Instead, we rely on CNI plugin that is running in the root
> namespace context to configure the VF interface as needed. (I've
> contributed custom MAC support to SR-IOV CNI plugin very recently.)
> Ihar

libvirt-users mailing list