Re: [PATCH v2] sched/fair: sanitize vruntime of entity being migrated

[Vincent Guittot <vincent.guittot@linaro.org>]

Le jeudi 09 mars 2023 à 16:14:38 (+0100), Vincent Guittot a écrit :
> On Thu, 9 Mar 2023 at 15:37, Peter Zijlstra <peterz@infradead.org> wrote:
> >
> > On Thu, Mar 09, 2023 at 03:28:25PM +0100, Peter Zijlstra wrote:
> > > On Thu, Mar 09, 2023 at 02:34:05PM +0100, Vincent Guittot wrote:
> > >
> > > > Then, even if we don't clear exec_start before migrating and keep
> > > > current value to be used in place_entity on the new cpu, we can't
> > > > compare the rq_clock_task(rq_of(cfs_rq)) of 2 different rqs AFAICT
> > >
> > > Blergh -- indeed, irq and steal time can skew them between CPUs :/
> > > I suppose we can fudge that... wait_start (which is basically what we're
> > > making it do) also does that IIRC.
> > >
> > > I really dislike having this placement muck spreadout like proposed.
> >
> > Also, I think we might be over-engineering this, we don't care about
> > accuracy at all, all we really care about is 'long-time'.
> 
> you mean taking the patch 1/2 that you mentioned here to add a
> migrated field:
> https://lore.kernel.org/all/68832dfbb60fda030540b5f4e39c5801942689b1.1648228023.git.tim.c.chen@linux.intel.com/T/#ma5637eb8010f3f4a4abff778af8db705429d003b
> 
> And assume that the divergence between the rq_clock_task() can be ignored ?
> 
> That could probably work but we need to replace the (60LL *
> NSEC_PER_SEC) by ((1ULL << 63) / NICE_0_LOAD) because 60sec divergence
> would not be unrealistic.
> and a comment to explain why it's acceptable

Zhang,

Could you try the patch below ?
This is a rebase/merge/update of:
-patch 1/2 above and 
-https://lore.kernel.org/lkml/20230209193107.1432770-1-rkagan@amazon.de/

The proposal accepts a divergence of up to 52 days between the 2 rqs.

If this work, we will prepare a proper patch

diff --git a/include/linux/sched.h b/include/linux/sched.h
index 63d242164b1a..cb8af0a137f7 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -550,6 +550,7 @@ struct sched_entity {
        struct rb_node                  run_node;
        struct list_head                group_node;
        unsigned int                    on_rq;
+       unsigned int                    migrated;

        u64                             exec_start;
        u64                             sum_exec_runtime;
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 7a1b1f855b96..36acd9598b40 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -1057,6 +1057,7 @@ update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
        /*
         * We are starting a new run period:
         */
+       se->migrated = 0;
        se->exec_start = rq_clock_task(rq_of(cfs_rq));
 }

@@ -4684,13 +4685,23 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)

        /*
         * Pull vruntime of the entity being placed to the base level of
-        * cfs_rq, to prevent boosting it if placed backwards.  If the entity
-        * slept for a long time, don't even try to compare its vruntime with
-        * the base as it may be too far off and the comparison may get
-        * inversed due to s64 overflow.
+        * cfs_rq, to prevent boosting it if placed backwards.
+        * However, min_vruntime can advance much faster than real time, with
+        * the exterme being when an entity with the minimal weight always runs
+        * on the cfs_rq. If the new entity slept for long, its vruntime
+        * difference from min_vruntime may overflow s64 and their comparison
+        * may get inversed, so ignore the entity's original vruntime in that
+        * case.
+        * The maximal vruntime speedup is given by the ratio of normal to
+        * minimal weight: NICE_0_LOAD / MIN_SHARES, so cutting off on the
+        * sleep time of 2^63 / NICE_0_LOAD should be safe.
+        * When placing a migrated waking entity, its exec_start has been set
+        * from a different rq. In order to take into account a possible
+        * divergence between new and prev rq's clocks task because of irq and
+        * stolen time, we take an additional margin.
         */
        sleep_time = rq_clock_task(rq_of(cfs_rq)) - se->exec_start;
-       if ((s64)sleep_time > 60LL * NSEC_PER_SEC)
+       if ((s64)sleep_time > (1ULL << 63) / NICE_0_LOAD / 2)
                se->vruntime = vruntime;
        else
                se->vruntime = max_vruntime(se->vruntime, vruntime);
@@ -7658,7 +7669,7 @@ static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
        se->avg.last_update_time = 0;

        /* We have migrated, no longer consider this task hot */
-       se->exec_start = 0;
+       se->migrated = 1;

        update_scan_period(p, new_cpu);
 }
@@ -8344,6 +8355,9 @@ static int task_hot(struct task_struct *p, struct lb_env *env)
        if (sysctl_sched_migration_cost == 0)
                return 0;

+       if (p->se.migrated)
+               return 0;
+
        delta = rq_clock_task(env->src_rq) - p->se.exec_start;

        return delta < (s64)sysctl_sched_migration_cost;



> 
> 
> >
> >