[libvirt] [PATCH 0/4] qemu: Honor memory mode='strict'

[Michal Privoznik]

On 4/10/19 12:35 AM, Daniel Henrique Barboza wrote:
> Hi,
> 
> On 4/9/19 11:10 AM, Michal Privoznik wrote:
>> If there's a domain configured as:
>>
>>    <currentMemory unit='MiB'>4096</currentMemory>
>>    <numatune>
>>      <memory mode='strict' nodeset='1'/>
>>    </numatune>
>>
>> but there is not enough memory on NUMA node 1 the domain will start
>> successfully because we allow it to allocate memory from other nodes.
>> This is a result of some previous fix (v1.2.7-rc1~91). However, I've
>> tested my fix successfully on a NUMA machine with F29 and recent kernel.
>> So the kernel bug I'm mentioning in 4/4 is probably fixed then and we
>> can drop the workaround.
> 
> I've tested out of curiosity your patch set in a Power8 system to see
> if I could spot a difference, but in my case it didn't change the behavior.
> 
> I've tried a guest with the following numatune:
> 
> 
>    <memory unit='KiB'>67108864</memory>
>    <currentMemory unit='KiB'>67108864</currentMemory>
>    <vcpu placement='static' current='4'>16</vcpu>
>    <numatune>
>      <memory mode='strict' nodeset='0'/>
>    </numatune>
> 
> This is the numa setup of the host:
> 
> $ numactl -H
> available: 4 nodes (0-1,16-17)
> node 0 cpus: 0 8 16 24 32 40
> node 0 size: 32606 MB
> node 0 free: 24125 MB
> node 1 cpus: 48 56 64 72 80 88
> node 1 size: 32704 MB
> node 1 free: 27657 MB
> node 16 cpus: 96 104 112 120 128 136
> node 16 size: 32704 MB
> node 16 free: 25455 MB
> node 17 cpus: 144 152 160 168 176 184
> node 17 size: 32565 MB
> node 17 free: 30030 MB
> node distances:
> node   0   1  16  17
>    0:  10  20  40  40
>    1:  20  10  40  40
>   16:  40  40  10  20
>   17:  40  40  20  10
> 
> 
> If I understood it right, the patches removed the capability to allocate
> memory from different numa nodes with the 'strict' setting, making the
> guest failing to launch if the numa node does not have enough memory.
> Unless I am getting something wrong, this guest shouldn't launch after
> applying this patch (node 0 does not have 64Gb available). But the guest
> is launching as if nothing changed.
> 
> 
> I'll dig it further if I have the chance. I'm just curious if this is 
> something
> that works differently with pseries guests.

Hey,

firstly thanks for testing this! And yes, it shows flaw in my patches. 
Thing is, my patches set up emulator/cpuset.mems but at the time qemu is 
doing its allocation it is still living under top level CGroup (which is 
left untouched). Will post v2! Thanks.

Michal