On Tue, Jan 26, 2016 at 1:51 PM, Michal Privoznik
<mprivozn(a)redhat.com>
wrote:
> On 26.01.2016 14:35, Andrei Perietanu wrote:
>> On Tue, Jan 26, 2016 at 12:39 PM, Michal Privoznik <mprivozn(a)redhat.com>
>> wrote:
>>
>>> On 26.01.2016 12:30, Andrei Perietanu wrote:
>>>> Hi all,
>>>>
>>>> I am running KVM on a 3.18 kernel. The system runs and Atom processor
>>> with
>>>> 2Gb RAM.
>>>>
>>>> Using KVM you obviously can over allocate your resources: say you have
> 4
>>>> guests each configured with 1GB ram. Running all four at the same time,
>>>> depending on the workload, can crash the system - I get a kernel trace
>>> when
>>>> this happens.
>>>>
>>>> But let's consider a simpler case: one guest with 1.5 Gb RAM, ubuntu
>>> 14.03.
>>>> During the installation the system will again crash.
>>>> The memory statistics (top or proc/meminfo) will show that the
> FreeMemory
>>>> goes down to 12Mb when this happens - which kind of makes sense
>>> considering
>>>> the host will require some RAM to run.
>>>>
>>>> But the question is: does libvirt offer any way to prevent this from
>>>> happening?
>>>>
>>>> Some way of not allowing the user to start a guest unless you have
> enough
>>>> free memory. I know how much ram each guest has configured but that is
>>> not
>>>> enough. I need to know how much the system has available, and just
>>> reading
>>>> the free memory statistic does not help much since that is only a
>>> snapshot
>>>> - when running a guest you can have 1gb free now, and 10 mb free 2 min
>>>> later.
>>>>
>>>> Any ideas?
>>>
>>> There is one option I see, use -mem-prealloc. Either you can passthrough
>>> it onto qemu commandline [1] or use locked memoryBacking [2]. I advocate
>>> for the latter though. Not only it will allocate all the memory at qemu
>>> startup it will also lock it so it won't get swapped off.
>>>
>>> Michal
>>>
>>> 1:
http://libvirt.org/drvqemu.html#qemucommand
>>> 2:
http://libvirt.org/formatdomain.html#elementsMemoryBacking
>>>
>>
>> I tried memoryBacking: I added this to the domain xml:
>> <memoryBacking>
>> <locked/>
>> </memoryBacking>
>>
>> And got an error when attempting to start the vm: memory locking no
>> supported by QEMU binary.
>>
>> Aside from that I don't really understand how this helps solve my issue -
>> if you don't mind going a bit into details I'd appreciate it.
>>
>
> Well, it will make qemu to allocate all its memory on the start. So
> either the allocation would be successful and domain will run or it
> won't and qemu will die immediately.
>
> On the other hand, your kernel is heavily broken too. Instead of
> crashing at OOM it should kill a process (usually the one that's
> consuming the most memory => qemu in this case).
>
> Michal
>
Sorry, I said crashing; I meant it spits out a kernel trace. You can still
use the system after it happens. I just want to prevent from happening in
the first place.
Well, then try the first one (passing -mem-prealloc onto qemu cmd line).
Also, you may want to turn off overcomit if you are unhappy with the
traces (turning it off will actually enable malloc to return NULL on OOM):
# echo 0 > /proc/sys/vm/overcommit_ratio
At this point I am gonna mention the obvious - don't give so much ram if
you don't have enough to give. For instance, give only 1GB, or use swap
so that unused tasks can be swapped out for a moment. Libvirt does not
actually have a mechanism to ensure you have enough memory on your host
to run a machine. It's host's kernel responsibility to deal with that.
Michal