[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-devel] [PATCH 5/5] spapr: drop DIV_ROUND_UP() from xics_max_se
From: |
David Gibson |
Subject: |
Re: [Qemu-devel] [PATCH 5/5] spapr: drop DIV_ROUND_UP() from xics_max_server_number() |
Date: |
Thu, 15 Feb 2018 15:08:18 +1100 |
User-agent: |
Mutt/1.9.2 (2017-12-15) |
On Wed, Feb 14, 2018 at 08:41:03PM +0100, Greg Kurz wrote:
> XICS needs to know the highest VCPU id that may be presented to the
> guest plus 1. Commit f303f117fec3 "spapr: ensure we have at least one
> XICS server" changed how the maximum is computed from:
>
> smp_cpus * kvmppc_smt_threads() / smp_threads
>
> to:
>
> DIV_ROUND_UP(smp_cpus * kvmppc_smt_threads(), smp_threads)
>
> This was done because at the time we could pass broken CPU topologies
> to the -smp command line options, such as threads=9,cpus=1. On a POWER8
> host this would give:
>
> 1 * 8 / 9 == 0 servers
>
> and cause QEMU to crash later during XICS setup.
>
> The formulat evolved a bit to accomodate CPU hot-plug and VSMT, but
> most important, stricter checks are performed on the CPU topology.
>
> With -smp threads=9,cpus=1:
>
> qemu-system-ppc64:
> cpu topology: sockets (1) * cores (1) * threads (9) > maxcpus (1)
>
> With -smp threads=9,maxcpus=1:
>
> qemu-system-ppc64: maxcpus must be equal to or greater than smp
>
> More generally, machine types with hotplug support (2.7 and up), no
> longer allow to set maxcpus or smp_cpus to a value that isnt't a
> multiple of smp_threads.
>
> With -smp threads=4,cpus=6:
>
> qemu-system-ppc64: smp_cpus (6) must be multiple of threads (4)
>
> With -smp threads=4,maxcpus=6:
>
> qemu-system-ppc64: max_cpus (6) must be multiple of threads (4)
>
> This means that the division is perfect and we don't need DIV_ROUND_UP(),
> and we could do a regular division:
>
> max_cpus * spapr->vsmt / smp_threads
>
> So this patch changes xics_max_server_number() to use the spapr_vcpu_id(),
> which works too since max_cpus is a multiple of smp_threads:
>
> (max_cpus / smp_threads ) * spapr->vsmt + max_cpus % smp_threads
>
> It breaks migration of pre-2.7 machine types with unusual CPU topologies,
> but I guess this is an acceptable trade-off.
No, not really. Weird topologies are still allowed on old machine
types for backwards compatibility, and we shouldn't break that. I
like the idea of consolidating this calculation, but we can't do it by
just breaking the older machines (at least not until they're formally
deprecated).
--
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
- [Qemu-devel] [PATCH 0/5] spapr: fix VCPU ids miscalculation, Greg Kurz, 2018/02/14
- [Qemu-devel] [PATCH 1/5] spapr: use spapr->vsmt to compute VCPU ids, Greg Kurz, 2018/02/14
- [Qemu-devel] [PATCH 2/5] spapr: move VCPU calculation to core machine code, Greg Kurz, 2018/02/14
- [Qemu-devel] [PATCH 3/5] spapr: rename spapr_vcpu_id() to spapr_get_vcpu_id(), Greg Kurz, 2018/02/14
- [Qemu-devel] [PATCH 4/5] spapr: consolidate the VCPU id numbering logic in a single place, Greg Kurz, 2018/02/14
- [Qemu-devel] [PATCH 5/5] spapr: drop DIV_ROUND_UP() from xics_max_server_number(), Greg Kurz, 2018/02/14
- Re: [Qemu-devel] [PATCH 5/5] spapr: drop DIV_ROUND_UP() from xics_max_server_number(),
David Gibson <=