[Libguestfs] [libvirt] Quantifying libvirt errors in launching the libguestfs appliance

[Cole Robinson]

On 01/14/2016 05:12 AM, Daniel P. Berrange wrote:
> On Thu, Jan 14, 2016 at 10:51:47AM +0100, Jiri Denemark wrote:
>> On Wed, Jan 13, 2016 at 16:25:14 +0100, Martin Kletzander wrote:
>>> On Wed, Jan 13, 2016 at 10:18:42AM +0000, Richard W.M. Jones wrote:
>>>> As people may know, we frequently encounter errors caused by libvirt
>>>> when running the libguestfs appliance.
>>>>
>>>> I wanted to find out exactly how frequently these happen and classify
>>>> the errors, so I ran the 'virt-df' tool overnight 1700 times.  This
>>>> tool runs several parallel qemu:///session libvirt connections both
>>>> creating a short-lived appliance guest.
>>>>
>>>> Note that I have added Cole's patch to fix https://bugzilla.redhat.com/1271183
>>>> "XML-RPC error : Cannot write data: Transport endpoint is not connected"
>>>>
>>>> Results:
>>>>
>>>> The test failed 538 times (32% of the time), which is pretty dismal.
>>>> To be fair, virt-df is aggressive about how it launches parallel
>>>> libvirt connections.  Most other virt-* tools use only a single
>>>> libvirt connection and are consequently more reliable.
>>>>
>>>> Of the failures, 518 (96%) were of the form:
>>>>
>>>>  process exited while connecting to monitor: qemu: could not load kernel '/home/rjones/d/libguestfs/tmp/.guestfs-1000/appliance.d/kernel': Permission denied
>>>>
>>>> which is https://bugzilla.redhat.com/921135 or maybe
>>>> https://bugzilla.redhat.com/1269975.  It's not clear to me if these
>>>> bugs have different causes, but if they do then potentially we're
>>>> seeing a mix of both since my test has no way to distinguish them.
>>>>
>>>
>>> It looks to me as the same problem.  And as the same problem we were
>>> talking about bunch of time and, apparently, didn't get to a conclusion.
>>>
>>> For each of the kernels, libvirt labels them (with both DAC and selinux
>>> labels), then proceeds to launching qemu.  If this is done parallel, the
>>> race is pretty obvious.  Could you remind me why you couldn't use
>>> <seclabel model='none'/> or <seclabel relabel='no'/> or something that
>>> would mitigate this?  If we cannot use this, then we need to implement
>>> the <seclabel/> element for kernel and initrd.
>>
>> Hmm, can't we just label kernel and initrd files the same way we label
>> <shareable/> disk images, i.e., non-exclusive label so that all QEMU
>> process can access them and avoid removing the label once a domain
>> disappears?
> 
> We actually should treat it in the same way as <readonly/> disks,
> and give it a shared read-only label. And indeed we *do* that.
> 
> The difference comes in the restore step - where we blow away the
> readonly label and put it back to the original. For disks we never
> restore readonly/shared labels, but for kernels we do. If we just
> kill the restore step for kernels too, we should be fine AFAICT.
> 

Indeed I forgot we don't restore labels on readonly/shareable disks..
certainly kernel/initrd should match that.

- Cole