| Message ID | 20231024214625.6483-3-frederic@kernel.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | rcu: Fix PF_IDLE related issues v2 | expand |
On Tue, Oct 24, 2023 at 11:46:23PM +0200, Frederic Weisbecker wrote: > +/* Check for quiescent states since the pregp's synchronize_rcu() */ > +static bool rcu_tasks_is_holdout(struct task_struct *t) > +{ > + int cpu; > + > + /* Has the task been seen voluntarily sleeping? */ > + if (!READ_ONCE(t->on_rq)) > + return false; > + > + cpu = task_cpu(t); > + > + /* > + * Idle tasks within the idle loop or offline CPUs are RCU-tasks > + * quiescent states. But CPU boot code performed by the idle task > + * isn't a quiescent state. > + */ > + if (t == idle_task(cpu)) { > + if (is_idle_task(t)) > + return false; > + > + if (!rcu_cpu_online(cpu)) > + return false; > + } Hmm, why is this guarded by t == idle_task() ? Notably, there is the idle-injection thing that uses FIFO tasks to run 'idle', see play_idle_precise(). This will (temporarily) get PF_IDLE on tasks that are not idle_task(). > + > + return true; > +} > + > /* Per-task initial processing. */ > static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop) > { > - if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) { > + if (t != current && rcu_tasks_is_holdout(t)) {
Le Wed, Oct 25, 2023 at 10:40:08AM +0200, Peter Zijlstra a écrit : > On Tue, Oct 24, 2023 at 11:46:23PM +0200, Frederic Weisbecker wrote: > > > +/* Check for quiescent states since the pregp's synchronize_rcu() */ > > +static bool rcu_tasks_is_holdout(struct task_struct *t) > > +{ > > + int cpu; > > + > > + /* Has the task been seen voluntarily sleeping? */ > > + if (!READ_ONCE(t->on_rq)) > > + return false; > > + > > + cpu = task_cpu(t); > > + > > + /* > > + * Idle tasks within the idle loop or offline CPUs are RCU-tasks > > + * quiescent states. But CPU boot code performed by the idle task > > + * isn't a quiescent state. > > + */ > > + if (t == idle_task(cpu)) { > > + if (is_idle_task(t)) > > + return false; > > + > > + if (!rcu_cpu_online(cpu)) > > + return false; > > + } > > Hmm, why is this guarded by t == idle_task() ? > > Notably, there is the idle-injection thing that uses FIFO tasks to run > 'idle', see play_idle_precise(). This will (temporarily) get PF_IDLE on > tasks that are not idle_task(). Ah good point. So indeed the is_idle_task() test doesn't musn't be guarded by t == idle_task(cpu). But rcu_cpu_online() has to, otherwise if it's not an idle task, there is a risk that the task gets migrated out by the time we observe the old CPU offline. Thanks. > > > + > > + return true; > > +} > > + > > /* Per-task initial processing. */ > > static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop) > > { > > - if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) { > > + if (t != current && rcu_tasks_is_holdout(t)) { > >
> > Le Wed, Oct 25, 2023 at 10:40:08AM +0200, Peter Zijlstra a écrit : > > On Tue, Oct 24, 2023 at 11:46:23PM +0200, Frederic Weisbecker wrote: > > > > > +/* Check for quiescent states since the pregp's synchronize_rcu() */ > > > +static bool rcu_tasks_is_holdout(struct task_struct *t) > > > +{ > > > + int cpu; > > > + > > > + /* Has the task been seen voluntarily sleeping? */ > > > + if (!READ_ONCE(t->on_rq)) > > > + return false; > > > + > > > + cpu = task_cpu(t); > > > + > > > + /* > > > + * Idle tasks within the idle loop or offline CPUs are RCU-tasks > > > + * quiescent states. But CPU boot code performed by the idle task > > > + * isn't a quiescent state. > > > + */ > > > + if (t == idle_task(cpu)) { > > > + if (is_idle_task(t)) > > > + return false; > > > + > > > + if (!rcu_cpu_online(cpu)) > > > + return false; > > > + } > > > > Hmm, why is this guarded by t == idle_task() ? > > > > Notably, there is the idle-injection thing that uses FIFO tasks to run > > 'idle', see play_idle_precise(). This will (temporarily) get PF_IDLE on > > tasks that are not idle_task(). > > Ah good point. So indeed the is_idle_task() test doesn't musn't be > guarded by t == idle_task(cpu). But rcu_cpu_online() has to, otherwise > if it's not an idle task, there is a risk that the task gets migrated out > by the time we observe the old CPU offline. > If a fifo-tasks use play_idle_precise() to run idle and invoke do_idle(), may cause rcu-tasks to falsely report a rcu-tasks QS , when rcu_is_cpu_rrupt_from_idle() return true in rcu_sched_clock_irq(), so should we also add a check for "current == idle_task(task_cpu(current))" in the rcu_is_cpu_rrupt_from_idle() ? Thanks Zqiang > Thanks. > > > > > > + > > > + return true; > > > +} > > > + > > > /* Per-task initial processing. */ > > > static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop) > > > { > > > - if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) { > > > + if (t != current && rcu_tasks_is_holdout(t)) { > > > >
On Thu, Oct 26, 2023 at 08:15:33PM +0800, Z qiang wrote: > > > > Le Wed, Oct 25, 2023 at 10:40:08AM +0200, Peter Zijlstra a écrit : > > > On Tue, Oct 24, 2023 at 11:46:23PM +0200, Frederic Weisbecker wrote: > > > > > > > +/* Check for quiescent states since the pregp's synchronize_rcu() */ > > > > +static bool rcu_tasks_is_holdout(struct task_struct *t) > > > > +{ > > > > + int cpu; > > > > + > > > > + /* Has the task been seen voluntarily sleeping? */ > > > > + if (!READ_ONCE(t->on_rq)) > > > > + return false; > > > > + > > > > + cpu = task_cpu(t); > > > > + > > > > + /* > > > > + * Idle tasks within the idle loop or offline CPUs are RCU-tasks > > > > + * quiescent states. But CPU boot code performed by the idle task > > > > + * isn't a quiescent state. > > > > + */ > > > > + if (t == idle_task(cpu)) { > > > > + if (is_idle_task(t)) > > > > + return false; > > > > + > > > > + if (!rcu_cpu_online(cpu)) > > > > + return false; > > > > + } > > > > > > Hmm, why is this guarded by t == idle_task() ? > > > > > > Notably, there is the idle-injection thing that uses FIFO tasks to run > > > 'idle', see play_idle_precise(). This will (temporarily) get PF_IDLE on > > > tasks that are not idle_task(). > > > > Ah good point. So indeed the is_idle_task() test doesn't musn't be > > guarded by t == idle_task(cpu). But rcu_cpu_online() has to, otherwise > > if it's not an idle task, there is a risk that the task gets migrated out > > by the time we observe the old CPU offline. > > > > If a fifo-tasks use play_idle_precise() to run idle and invoke > do_idle(), may cause > rcu-tasks to falsely report a rcu-tasks QS Well, there can be a debate here: should we consider an idle injector as a real task that we must wait for a voluntary schedule or should we treat it just like an idle task? Having that whole idle task quiescent state in RCU-tasks is quite a strange semantic anyway. And in the long run, the purpose is to unify RCU-tasks and RCU-tasks-RUDE with relying on ct_dynticks for idle quiescent states. > , when rcu_is_cpu_rrupt_from_idle() > return true in rcu_sched_clock_irq(), so should we also add a check for > "current == idle_task(task_cpu(current))" in the rcu_is_cpu_rrupt_from_idle() > ? That looks fine OTOH. Whether idle injection or real idle, rcu_is_cpu_rrupt_from_idle() is always a quiescent state in real RCU. Because we know we have no RCU reader between ct_idle_enter() and ct_idle_exit(). Thanks.
diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h index bf5f178fe723..acf81efe5eff 100644 --- a/kernel/rcu/tasks.h +++ b/kernel/rcu/tasks.h @@ -895,10 +895,37 @@ static void rcu_tasks_pregp_step(struct list_head *hop) synchronize_rcu(); } +/* Check for quiescent states since the pregp's synchronize_rcu() */ +static bool rcu_tasks_is_holdout(struct task_struct *t) +{ + int cpu; + + /* Has the task been seen voluntarily sleeping? */ + if (!READ_ONCE(t->on_rq)) + return false; + + cpu = task_cpu(t); + + /* + * Idle tasks within the idle loop or offline CPUs are RCU-tasks + * quiescent states. But CPU boot code performed by the idle task + * isn't a quiescent state. + */ + if (t == idle_task(cpu)) { + if (is_idle_task(t)) + return false; + + if (!rcu_cpu_online(cpu)) + return false; + } + + return true; +} + /* Per-task initial processing. */ static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop) { - if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) { + if (t != current && rcu_tasks_is_holdout(t)) { get_task_struct(t); t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw); WRITE_ONCE(t->rcu_tasks_holdout, true); @@ -947,7 +974,7 @@ static void check_holdout_task(struct task_struct *t, if (!READ_ONCE(t->rcu_tasks_holdout) || t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) || - !READ_ONCE(t->on_rq) || + !rcu_tasks_is_holdout(t) || (IS_ENABLED(CONFIG_NO_HZ_FULL) && !is_idle_task(t) && READ_ONCE(t->rcu_tasks_idle_cpu) >= 0)) { WRITE_ONCE(t->rcu_tasks_holdout, false);
The commit: cff9b2332ab7 ("kernel/sched: Modify initial boot task idle setup") has changed the semantics of what is to be considered an idle task in such a way that CPU boot code preceding the actual idle loop is excluded from it. This has however introduced new potential RCU-tasks stalls when either: 1) Grace period is started before init/0 had a chance to set PF_IDLE, keeping it stuck in the holdout list until idle ever schedules. 2) Grace period is started when some possible CPUs have never been online, keeping their idle tasks stuck in the holdout list until the CPU ever boots up. 3) Similar to 1) but with secondary CPUs: Grace period is started concurrently with secondary CPU booting, putting its idle task in the holdout list because PF_IDLE isn't yet observed on it. It stays then stuck in the holdout list until that CPU ever schedules. The effect is mitigated here by the hotplug AP thread that must run to bring the CPU up. Fix this with handling the new semantics of PF_IDLE, keeping in mind that it may or may not be set on an idle task. Take advantage of that to strengthen the coverage of an RCU-tasks quiescent state within an idle task, excluding the CPU boot code from it. Only the code running within the idle loop is now a quiescent state, along with offline CPUs. Fixes: cff9b2332ab7 ("kernel/sched: Modify initial boot task idle setup") Suggested-by: Joel Fernandes <joel@joelfernandes.org> Suggested-by: Paul E . McKenney" <paulmck@kernel.org> Signed-off-by: Frederic Weisbecker <frederic@kernel.org> --- kernel/rcu/tasks.h | 31 +++++++++++++++++++++++++++++-- 1 file changed, 29 insertions(+), 2 deletions(-)