[libvirt] question about rdma migration

Roy Shterman roysh at mellanox.com
Tue Mar 22 13:16:49 UTC 2016


OK, I will describe it better,

I'm developing iSER transport option inside libiscsi, as part of that I'm planning to implement it also in QEMU block layer.

iSER RDMA components (QP, CQ, MR) need to lock some amount of memory (predictable amount),

So, needed memory to be locked is (num of libiscsi-iser devices in VM)*(constant amount per device),

For now, I'm using locked option in libvirt, although I don't really need to lock all VM memory.

Regards,
Roy

-----Original Message-----
From: Jiri Denemark [mailto:jdenemar at redhat.com] 
Sent: Tuesday, March 22, 2016 2:56 PM
To: Roy Shterman <roy.shterman at gmail.com>
Cc: libvir-list at redhat.com; Roy Shterman <roysh at mellanox.com>
Subject: Re: [libvirt] question about rdma migration

On Tue, Mar 22, 2016 at 14:21:52 +0200, Roy Shterman wrote:
> Correct me if I'm wrong but locked option is pinning all VM memory in 
> host RAM,
> 
> for example if I have a VM with 4G memory, and I want to run some QEMU 
> code which needs to pin 500M,
> 
> I will need to lock all 4G in host memory instead of locking only 500M.

So the question is which code wants to lock part of the memory, why, and if it's something that can be influenced by user.

For example, we know that if you ask for all memory to by locked, we need to set the limit. The same applies when RDMA migration is started.
On PPC we know some amount of memory will always need to be locked, we compute the amount and set the limit accordingly. We can't really expect user to have deep knowledge of QEMU and know what limit needs to be set when they use a specific device, QMP command, or whatever. So if the limit is something predictable and deterministic, we can automatically compute the amount of memory and use it when starting QEMU. Forcing users to set the limit when all memory needs to be locked is already bad enough that I don't think we should add a new option to explicitly set arbitrary lock limit.

Jirka




More information about the libvir-list mailing list