[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[Qemu-devel] [repost] KVM call minutes for 2016-02-16
From: |
Juan Quintela |
Subject: |
[Qemu-devel] [repost] KVM call minutes for 2016-02-16 |
Date: |
Tue, 23 Feb 2016 09:35:39 +0100 |
User-agent: |
Gnus/5.13 (Gnus v5.13) Emacs/24.5 (gnu/linux) |
Hi
[I mistyped qemu-devel@ address on first submission, sorry.]
Random ideas that I took for the talk, I hope that some of the active
participants can reply with a more coherent view.
State of the art (Andreas)
- Having very specific socket objects
- Dynamic allocation using new parameters
- Which level should we be doing the hotplug (granularity)
x86-> socket level
powerpc -> core
z390 -> thread level
Christian
- Low level device adding
- Initialize from the beginning all cpu objects
Hot-plugging would be to put them on-line
- how cpu came into being
- how are we going to add in x86 with device add
* x86 and sparc -cpu handling is special case
* not copy that on s390, better example in ARM?
- Way to handle this?
* save the feature string internally, for reusing with new cpu
* how best handle, it can depend of cpu model and board
* how to make libvirt level not crazy?
* should we hide the cpu add inside the machine implementation
* having an array of empty sockets where to plug new cpus
* only compatible types can be set on link property (not sure how well are
enforced)
* set the link property of the machine to the core object just created
* powerpc cpu object is realized hook should check that it is done
is a hot plug operation to automate this
* powerpc only has cores and threads
* we can have a property in the core object to know where it goes
* if each core has two threads, when we add a new core, we always add 2
threads
* cpu_add have a type={core,thread,whatever} what is the problem with it?
* pre-created id's for cpu_add?
* do different implementations and then try to find commonalities.
* ARM can need to be able to add different cpu models
* how to specify the topology
* cpu_add has always thought to be an interim solution until we get the real
one
* problems with migration, specially with cpu_del
* get cpu_add on x86
* they don't support unplug
* they don't support topology
* realization matters about where we put the things, in child or parent object
Later, JUan.
[Prev in Thread] |
Current Thread |
[Next in Thread] |
- [Qemu-devel] [repost] KVM call minutes for 2016-02-16,
Juan Quintela <=