Re: [External] Re: [PATCH 2/2] sched/core: Avoid double calling update_rq_clock()

[Vincent Guittot]

Le mardi 02 mai 2023 à 18:22:56 (+0800), Hao Jia a écrit :
> 
> 
> On 2023/4/21 Vincent Guittot wrote:
> > On Mon, 10 Apr 2023 at 10:12, Hao Jia <jiahao.os@xxxxxxxxxxxxx> wrote:
> > > 
> > > There are some double rq clock update warnings are triggered.
> > > ------------[ cut here ]------------
> > > rq->clock_update_flags & RQCF_UPDATED
> > > WARNING: CPU: 17 PID: 138 at kernel/sched/core.c:741
> > > update_rq_clock+0xaf/0x180
> > > Call Trace:
> > >   <TASK>
> > >   __balance_push_cpu_stop+0x146/0x180
> > >   ? migration_cpu_stop+0x2a0/0x2a0
> > >   cpu_stopper_thread+0xa3/0x140
> > >   smpboot_thread_fn+0x14f/0x210
> > >   ? sort_range+0x20/0x20
> > >   kthread+0xe6/0x110
> > >   ? kthread_complete_and_exit+0x20/0x20
> > >   ret_from_fork+0x1f/0x30
> > > 
> > > ------------[ cut here ]------------
> > > rq->clock_update_flags & RQCF_UPDATED
> > > WARNING: CPU: 54 PID: 0 at kernel/sched/core.c:741
> > > update_rq_clock+0xaf/0x180
> > > Call Trace:
> > >   <TASK>
> > >   unthrottle_cfs_rq+0x4b/0x300
> > >   __cfsb_csd_unthrottle+0xe0/0x100
> > >   __flush_smp_call_function_queue+0xaf/0x1d0
> > >   flush_smp_call_function_queue+0x49/0x90
> > >   do_idle+0x17c/0x270
> > >   cpu_startup_entry+0x19/0x20
> > >   start_secondary+0xfa/0x120
> > >   secondary_startup_64_no_verify+0xce/0xdb
> > > 
> > > ------------[ cut here ]------------
> > > rq->clock_update_flags & RQCF_UPDATED
> > > WARNING: CPU: 0 PID: 3323 at kernel/sched/core.c:741
> > > update_rq_clock+0xaf/0x180
> > > Call Trace:
> > >   <TASK>
> > >   unthrottle_cfs_rq+0x4b/0x300
> > >   rq_offline_fair+0x89/0x90
> > >   set_rq_offline.part.118+0x28/0x60
> > 
> > So this is generated by patch 1, isn't it ?
> 
> Sorry for the late reply, I just got back from a long term.
> 
> IIRC, this is not generated by patch1.

yeah, that's in the cfs loop

> 
> In the unthrottle_offline_cfs_rqs() function, we traverse task_groups
> through list_for_each_entry_rcu, so unthrottle_cfs_rq() may be called
> multiple times, resulting in multiple updates to the rq clock.
> 
> Thanks,
> Hao
> > 
> > >   rq_attach_root+0xc4/0xd0
> > >   cpu_attach_domain+0x3dc/0x7f0
> > >   partition_sched_domains_locked+0x2a5/0x3c0
> > >   rebuild_sched_domains_locked+0x477/0x830
> > >   rebuild_sched_domains+0x1b/0x30
> > >   cpuset_hotplug_workfn+0x2ca/0xc90
> > >   ? balance_push+0x56/0xf0
> > >   ? _raw_spin_unlock+0x15/0x30
> > >   ? finish_task_switch+0x98/0x2f0
> > >   ? __switch_to+0x291/0x410
> > >   ? __schedule+0x65e/0x1310
> > >   process_one_work+0x1bc/0x3d0
> > >   worker_thread+0x4c/0x380
> > >   ? preempt_count_add+0x92/0xa0
> > >   ? rescuer_thread+0x310/0x310
> > >   kthread+0xe6/0x110
> > >   ? kthread_complete_and_exit+0x20/0x20
> > >   ret_from_fork+0x1f/0x30
> > > 
> > > For the __balance_push_cpu_stop() case, we remove update_rq_clock() from
> > > the __migrate_task() function to avoid double updating the rq clock.
> > > And in order to avoid missing rq clock update, add update_rq_clock()
> > > call before migration_cpu_stop() calls __migrate_task().

Can we do the opposite ?
AFAICT, update_rq_clock() in __balance_push_cpu_stop() is only there for
__migrate_task(). I prefer to keep the update_rq_clock() as close as possible
to the user

> > > 
> > > This also works for unthrottle_cfs_rq(), so we also removed
> > > update_rq_clock() from the unthrottle_cfs_rq() function to avoid
> > > warnings caused by calling it multiple times, such as
> > > __cfsb_csd_unthrottle() and unthrottle_offline_cfs_rqs(). and

This happens with the for loop added by
commit: 8ad075c2eb1f ("sched: Async unthrottling for cfs bandwidth")

> > > in order to avoid missing rq clock update, we correspondingly add
> > > update_rq_clock() calls before unthrottle_cfs_rq() runs.

These are special cases that happen because of the for_each.
As said above, I would prefer keeping update_rq_clock close the their user

could we use something similar to rq_clock_skip_update() for those list ?

Something like the below:

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 898fa3bc2765..7495b6fb229d 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -9386,8 +9386,6 @@ static int __balance_push_cpu_stop(void *arg)
        raw_spin_lock_irq(&p->pi_lock);
        rq_lock(rq, &rf);

-       update_rq_clock(rq);
-
        if (task_rq(p) == rq && task_on_rq_queued(p)) {
                cpu = select_fallback_rq(rq->cpu, p);
                rq = __migrate_task(rq, &rf, p, cpu);
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 503b7617d7e1..a1d47d907360 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -5518,6 +5518,12 @@ static void __cfsb_csd_unthrottle(void *arg)
        struct rq_flags rf;

        rq_lock(rq, &rf);
+       /*
+        * Iterating over the list can trigger several call to update_rq_clock.
+        * Do it once and skip the potential next ones.
+        */
+       update_rq_clock(rq);
+       rq_clock_loop_update(rq);

        /*
         * Since we hold rq lock we're safe from concurrent manipulation of
@@ -6058,6 +6064,13 @@ static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)

        lockdep_assert_rq_held(rq);

+       /*
+        * Iterating over the list can trigger several call to update_rq_clock.
+        * Do it once and skip the potential next ones.
+        */
+       update_rq_clock(rq);
+       rq_clock_loop_update(rq);
+
        rcu_read_lock();
        list_for_each_entry_rcu(tg, &task_groups, list) {
                struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 66bacd50d381..681924367891 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -1536,6 +1536,19 @@ static inline void rq_clock_skip_update(struct rq *rq)
        rq->clock_update_flags |= RQCF_REQ_SKIP;
 }

+/*
+ * During cpu offlining and rq wide unthrottling, we can trigger
+ * an update_rq_clock() for several cfs and rt runqueues (Typically
+ * when using list_for_each_entry_*)
+ * rq_clock_loop_update() can be called after updating the clock once
+ * and before iterating over the list to prevent multiple update.
+ */
+static inline void rq_clock_loop_update(struct rq *rq)
+{
+       lockdep_assert_rq_held(rq);
+       rq->clock_update_flags |= RQCF_ACT_SKIP;
+}
+
 /*
  * See rt task throttling, which is the only time a skip
  * request is canceled.


> > > 
> > > Note that the rq clock has been updated before the set_rq_offline()
> > > function runs, so we don't need to add update_rq_clock() call in
> > > unthrottle_offline_cfs_rqs().
> > > 
> 
> 
>