Thread (29 messages) 29 messages, 4 authors, 2021-06-03

Re: [PATCH v5 2/3] sched/topology: Rework CPU capacity asymmetry detection

From: Beata Michalska <hidden>
Date: 2021-06-02 19:48:20
Also in: lkml
Subsystem: scheduler, the rest · Maintainers: Ingo Molnar, Peter Zijlstra, Juri Lelli, Vincent Guittot, Linus Torvalds

On Wed, Jun 02, 2021 at 07:17:12PM +0200, Dietmar Eggemann wrote:
quoted hunk ↗ jump to hunk
On 27/05/2021 19:07, Beata Michalska wrote:
quoted
On Thu, May 27, 2021 at 05:08:42PM +0200, Dietmar Eggemann wrote:
quoted
On 26/05/2021 23:40, Beata Michalska wrote:
quoted
On Wed, May 26, 2021 at 08:17:41PM +0200, Dietmar Eggemann wrote:
quoted
On 26/05/2021 14:51, Beata Michalska wrote:
quoted
On Wed, May 26, 2021 at 01:15:46PM +0100, Beata Michalska wrote:
quoted
On Wed, May 26, 2021 at 11:52:25AM +0200, Dietmar Eggemann wrote:
quoted
On 25/05/2021 12:29, Beata Michalska wrote:
quoted
On Tue, May 25, 2021 at 10:53:07AM +0100, Valentin Schneider wrote:
quoted
On 24/05/21 23:55, Beata Michalska wrote:
quoted
On Mon, May 24, 2021 at 07:01:04PM +0100, Valentin Schneider wrote:
quoted
On 24/05/21 11:16, Beata Michalska wrote:
[...]
quoted
quoted
quoted
We could possibly add a warning (like in EAS) if the asymmetry is detected
for SMT which would give some indication that there is smth ... wrong ?
Maybe, in case you find an easy way to detect this.

But the issue already exists today. Not with the topology mentioned
above but in case we slightly change it to:

  cpus = { ([446 1024] [871 1024] [446 1024] ) ([1024 1024]) }
                                       ^^^^
so that we have a 1024 CPU in the lowest sd for each CPU, we would get
SD_ASYM_CPUCAPACITY on SMT.
The asymmetry capacity flags are being set on a sched domain level, so
we could use the SD_SHARE_CPUCAPACITY|SD_SHARE_PKG_RESOURCES (cpu_smt_flags)
flags to determine if having asymmetry is valid or not ? If this is enough 
this could be handled by the classify function?
Or maybe something directly in sd_init(), like the WARN_ONCE() which triggers
if somebody wants to sneak in a ~topology flag via a
sched_domain_topology_level table? 

IMHO checking `SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY` will be sufficient
here.
diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
index 62d412013df8..77b73abbb9a4 100644
--- a/kernel/sched/topology.c
+++ b/kernel/sched/topology.c
@@ -1561,6 +1561,11 @@ sd_init(struct sched_domain_topology_level *tl,
        sd_id = cpumask_first(sched_domain_span(sd));
 
        sd->flags |= asym_cpu_capacity_classify(sd, cpu_map);
+
+       WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
+                 (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
+                     "CPU capacity asymmetry not supported on SMT\n");
+
        /*
         * Convert topological properties into behaviour.
         */
In case we can agree on something simple here I guess you can incorporate it into v7.
So what I have done is :
diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
index 77e6f79235ad..ec4ae225687e 100644
--- a/kernel/sched/topology.c
+++ b/kernel/sched/topology.c
@@ -1324,6 +1324,7 @@ asym_cpu_capacity_classify(struct sched_domain *sd,
        if (!asym_cap_miss)
                sd_asym_flags |= SD_ASYM_CPUCAPACITY_FULL;
 
+       WARN_ONCE(cpu_smt_flags() & sd->flags, "Detected CPU capacity asymmetry on SMT level");
 leave:
        return sd_asym_flags;
 }
Comment can be adjusted.
This would sit in the classify function to nicely wrap asymmetry bits in one
place. What do you think ?

---
BR
B.
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help