[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-ppc] [PATCH RFC] spapr: disintricate core-id from DT semantics
From: |
David Gibson |
Subject: |
Re: [Qemu-ppc] [PATCH RFC] spapr: disintricate core-id from DT semantics |
Date: |
Mon, 25 Jul 2016 16:01:31 +1000 |
User-agent: |
Mutt/1.6.2 (2016-07-01) |
On Fri, Jul 22, 2016 at 01:10:36PM +0200, Greg Kurz wrote:
> The goal of this patch is to have a stable core-id which does not depend
> on any DT related semantics, which involve non-obvious computations on
> modern PowerPC server cpus.
>
> With this patch, the DT core id is computed on-demand as:
>
> (core-id / smp_threads) * smt
>
> where smt is the number of threads per core in the host.
>
> This formula should be consolidated in a helper since it is needed in
> several places.
It's a little odd you node this but don't do so.
> Other uses for core-id includes: compute a stable cpu_index (which
> allows random order hotplug/unplug without breaking migration) and
> NUMA.
>
> Signed-off-by: Greg Kurz <address@hidden>
> ---
>
> It was first suggested here:
>
> https://lists.gnu.org/archive/html/qemu-devel/2016-07/msg01727.html
>
> and as option 1) in the following discussion on IRC:
>
> <dwg> imammedo, basically the options are: 1) change core-ids to be
> 0, 1, .. n and compute cpu_index as core_id * threads + thread#, or
> 2) leave core-ids as they are and calculate cpu_index as
> core-id / smt * threads + thread#
>
> It is based on David's ppc-for-2.7 branch (commit bb6268f35f457).
>
> It is lightly tested but I could at least do in-order
> hotplug/unplug.
I think this is basically the right approach, and I've applied to
ppc-for-2.7. Here's my plan for what to do about all this id stuff:
1. Merge this to ppc-for-2.7 (done)
2. Assuming there are now show-stoppers in testing, send a pull
request tomorrow
3. Once this is merged, try to get Igor's series (or a respin of it)
in ASAP.
I'm hoping that will give us good-enough hotplug for 2.7.
In the 2.8 timeframe, I want to:
4. Disconnect KVM vcpu ID from dt_id, instead calculate it from (now
stable) cpu_index
5. Remove dt_id as a cpu field - instead just compute DT ids from
the (now stable) cpu_index when we build the DT
6. (for new machine type versions) Change DT ID assignment, so it
no longer depends on kvmppc_smt_threads(). The current scheme
means that migration between hosts with different native SMT
values won't work, which is unfortunate. I suspect there may be
other problems with any real situation where that's the case,
but nontheless it's a silly restriction.
Nice to have but bigger scope things for 2.8:
7. Update archs to they *all* call cpu_exec_init() / cpu_exec_exit()
at realize / unrealize time instead of init / finalize time.
8. Update all archs and machines to use stable cpu_index
--
David Gibson | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_
| _way_ _around_!
http://www.ozlabs.org/~dgibson
signature.asc
Description: PGP signature