[PATCH 4/5] qemu: Prefer -numa cpu over -numa node,cpus=

[Igor Mammedov]

On Tue, 26 May 2020 17:31:09 +0200
Michal Privoznik <mprivozn at redhat.com> wrote:

> On 5/26/20 4:51 PM, Igor Mammedov wrote:
> > On Mon, 25 May 2020 10:05:08 +0200
> > Michal Privoznik <mprivozn at redhat.com> wrote:
> >   
> >>
> >> This is a problem. The domain XML that is provided can't be changed,
> >> mostly because mgmt apps construct it on the fly and then just pass it
> >> as a RO string to libvirt. While libvirt could create a separate cache,
> >> there has to be a better way.
> >>
> >> I mean, I can add some more code that once the guest is running
> >> preserves the mapping during migration. But that assumes a running QEMU.
> >> When starting a domain from scratch, is it acceptable it vCPU topology
> >> changes? I suspect it is not.  
> > I'm not sure I got you but
> > vCPU topology isn't changnig but when starting QEMU, user has to map
> > 'concrete vCPUs' to spencific numa nodes. The issue here is that
> > to specify concrete vCPUs user needs to get layout from QEMU first
> > as it's a function of target/machine/-smp and possibly cpu type.  
> 
> Assume the following config: 4 vCPUs (2 sockets, 2 cores, 1 thread 
> topology) and 2 NUMA nodes and the following assignment to NUMA:
> 
> node 0: cpus=0-1
> node 1: cpus=2-3
> 
> With old libvirt & qemu (and assuming x86_64 - not EPYC), I assume the 
> following topology is going to be used:
> 
> node 0: socket=0,core=0,thread=0 (vCPU0)  socket=0,core=1,thread=0 (vCPU1)
> node 1: socket=1,core=0,thread=0 (vCPU2)  socket=1,core=1,thread=0 (vCPU3)
> 
> Now, user upgrades libvirt & qemu but doesn't change the config. And on 
> a fresh new start (no migration), they might get a different topology:
> 
> node 0: socket=0,core=0,thread=0 (vCPU0)  socket=1,core=0,thread=0 (vCPU1)
> node 1: socket=0,core=1,thread=0 (vCPU2)  socket=1,core=1,thread=0 (vCPU3)

that shouldn't happen at least for as long as machine version stays the same

> (This is a very trivial example that I am intentionally making look bad, 
> but the thing is, there are some CPUs with very weird vCPU -> 
> socket/core/thread mappings).
>
> The problem here is that with this new version it is libvirt who 
> configured the vCPU -> NUMA mapping (using -numa cpu). Why so wrong? 
> Well it had no way to ask qemu how it used to be. Okay, so we add an 
> interface to QEMU (say -preconfig + query-hotpluggable-cpus) which will 
> do the mapping and keep it there indefinitely. But if the interface is 
> already there (and "always" will be), I don't see need for the extra 
> step (libvirt asking QEMU for the old mapping).
with cpu_index users don't know what CPUs they assing where,
and in some cases (spapr) it doesn't really maps into board supported CPU model well.
We can add and keep cpu_index in query-hotpluggable-cpus to help with migration
for old machine types from old CLI to the new one, but otherwise cpu_index would
disapear from user visible inerface.
I'd like to drop duplicate code supporting ambiguose '-numa node,cpus' (and not always
properly working interface) and keep only single variant '-numa cpu=' to do numa mapping,
which uses CPU's topology properties to describe CPUs, and unifies it with
the way it's done with cpu hotplug.

> The problem here is not how to assign vCPUs to NUMA nodes, the problem 
> is how to translate vCPU IDs to socket=,core=,thread=.
if you are talking about libvirt's vCPU IDs, then it's separate issue
as it's user facing API, I think it should not rely on cpu_index.
Instead it should map vCPU IDs to ([socket,]core[,thread]) tuple
or maybe drop notion of vCPU IDs and expose ([socket,]core[,thread])
to users if they ask for numa aware config.

PS:
I'm curious how libvirt currently implements numa mapping and
how it's correlated with pinnig to host nodes?
Does it have any sort of code to calculate topology based on cpu_index
so it could properly assign vCPUs to nodes or all the pain of
assigning vCPU IDs to nodes is on the user shoulders?

> > that applies not only '-numa cpu' but also to -device cpufoo,
> > that's why query-hotpluggable-cpus was introduced to let
> > user get the list of possible CPUs (including topo properties needed to
> > create them) for a given set of CLI options.
> > 
> > If I recall right libvirt uses topo properies during cpu hotplug but
> > treats it mainly as opaqueue info so it could feed it back to QEMU.
> > 
> >   
> >>>> tries to avoid that as much as it can.
> >>>>     
> >>>>>
> >>>>> How to present it to libvirt user I'm not sure (give them that list perhaps
> >>>>> and let select from it???)  
> >>>>
> >>>> This is what I am trying to figure out in the cover letter. Maybe we
> >>>> need to let users configure the topology (well, vCPU id to [socket, die,
> >>>> core, thread] mapping), but then again, in my testing the guest ignored
> >>>> that and displayed different topology (true, I was testing with -cpu
> >>>> host, so maybe that's why).  
> >>> there is ongiong issue with EPYC VCPUs topology, but I otherwise it should work.
> >>> Just report bug to qemu-devel, if it's broken.
> >>>        
> >>>>     
> >>>>> But it's irrelevant, to the patch, magical IDs for socket/core/...whatever
> >>>>> should not be generated by libvirt anymore, but rather taken from QEMU for given
> >>>>> machine + -smp combination.  
> >>>>
> >>>> Taken when? We can do this for running machines, but not for freshly
> >>>> started ones, can we?  
> >>>
> >>> it can be used for freshly started as well,
> >>> QEMU -S -preconfig -M pc -smp ...
> >>> (QMP) query-hotpluggable-cpus
> >>> (QMP) set-numa-node ...
> >>> ...
> >>> (QMP) exit-preconfig
> >>> (QMP) other stuff libvirt does (like hot-plugging CPUs , ...)
> >>> (QMP) cont  
> >>
> >> I'm not sure this works. query-hotpluggable-cpus does not map vCPU ID
> >> <-> socket/core/thread, For '-smp 2,sockets=2,cores=1,threads=1' the
> >> 'query-hotpluggable-cpus' returns:
> >>
> >> {"return": [{"props": {"core-id": 0, "thread-id": 0, "socket-id": 1},
> >> "vcpus-count": 1, "type": "qemu64-x86_64-cpu"}, {"props": {"core-id": 0,
> >> "thread-id": 0, "socket-id": 0}, "vcpus-count": 1, "type":
> >> "qemu64-x86_64-cpu"}]}  
> > 
> > that's the list I was taling about, which is implicitly ordered by cpu_index  
> 
> Aha! So in this case it would be:
> 
> vCPU0 -> socket=1,core=0,thread=0
> vCPU1 -> socket=0,core=0,thread=0
> 
> But that doesn't feel right. Is the cpu_index increasing or decreasing 
> as I go through the array?
it's array with decreasing order and index in it currently == cpu_index for
present and possible CPUs. Content of array is immutable for given
-M/-smp combination, to keep migration working. We can try to add
x-cpu-index to cpu entries, so you won't have to rely on order to help with
migrating from old CLI (but only for old machine types where old CLI actually
worked worked).

> Also, how is this able to express holes? E.g. 
> there might be some CPUs that don't have linear topology, and for 
> instance while socket=0,core=0,thread=0 and socket=0,core=0,thread=2 
> exist, socket=0,core=0,thread=1 does not. How am I supposed to know that 
> by just looking at the array?
speaking of x86, QEMU curently does not implement topologies with holes
in [socket/core/thread] tuple but if it were it shouldn't matter as all
CPUs and their realations with each other are described within that array.

 
> >> And 'query-cpus' or 'query-cpus-fast' which map vCPU ID onto
> >> socket/core/thread are not allowed in preconfig state.  
> > these 2 commands apply to present cpu only, if I'm not mistaken.
> > query-hotpluggable-cpus shows not only present but also CPUs that
> > could be hotplugged with device_add or used with -device.  
> 
> Fair enough. I haven't looked into the code that much.
> 
> > 
> >   
> >> But if I take a step back, the whole point of deprecating -numa
> >> node,cpus= is that QEMU no longer wants to do vCPU ID <->
> >> socket/core/thread mapping because it's ambiguous. So it feels a bit
> >> weird to design a solution where libvirt would ask QEMU to provide the
> >> mapping only so that it can be configured back. Not only because of the
> >> extra step, but also because QEMU can't then remove the mapping anyway.
> >> I might be misunderstanding the issue though.  
> > if '-numa node,cpus' is removed, we no longer will be using cpu_index as
> > configuration interface with user, that would allow QEMU start pruning
> > it from HMP/QMP interfaces and then probably remove it internally.
> > (I haven't explored yet if we could get rid of it completely but
> > I'd expect migration stream would be the only reason to keep it intrenally).
> > 
> > I'm quite reluctant to add cpu_index to modern query-hotpluggable-cpus output,
> > since the whole goal is to get rid of the index, which don't actually work
> > with SPAPR where CPU entity is a core and threads are internal impl. detail
> > (while cpu_index has 1:1 mapping with threads).
> > 
> > However if it will let QEMU to drop '-numa node,cpus=', we can discuss
> > adding optional 'x-cpu-index' to query-hotpluggable-cpus, that will be available
> > for old machine types for the sole purpose to help libvirt map old CLI to new one.
> > New machines shouldn't care about index though, since they should be using
> > '-numa cpu'.  
> 
> The problem here is that so far, all that libvirt users see are vCPU 
> IDs. They use them to assign vCPUs to NUMA nodes. And in order to make 
> libvirt switch to the new command line it needs a way to map IDs to 
> socket=,core=,thread=. I will play more with the preconfig and let you know.

If libvirt's vCPU IDs are mirroring cpu_index, I'd say it shouldn't be doing
so, see Daniel's response
https://lists.gnu.org/archive/html/qemu-devel/2017-10/msg04369.html

FYI:
I didn't read through all the history of -preconfig patches but QEMU options
for topology aware (sane) numa configuration on the table were:
  1. -numa node,cpus[cpu_index]
      libvirt needs to duplicate internal QEMU algorithms that map cpu_index
      values to topology info and use it to map vCPUs to numa nodes
      (and keep in sync with QEMU as it's machine versioned moving target)
  2. -numa cpu CLI option,
      libvirt needs to duplicate internal QEMU algorithms that calculate
      target depended values for socket/core/thread ids. (basically it's
      the same as #1), the only difference is that CLI user interface is
      expressed in topology properties.
  3. when we discussed it in the past #2 wasn't going to fly as it still
     had tha same burden as #1 (duplicating code and keeping it in sync).
     so we ended up with runtime configuration (-preconfig) to avoid QEMU
     restart just for querying, where libvirt could get list of possible CPUs
     from QEMU instance and complete numa configuration on the fly (at least
     for the first time, results could be cached and re-used with -numa cpu).