lkml.org 
[lkml]   [2019]   [Dec]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH tip/core/rcu] Enable tick for nohz_full CPUs not responding to expedited GP
On Thu, Dec 05, 2019 at 08:41:34PM -0500, Joel Fernandes wrote:
> On Wed, Dec 04, 2019 at 03:25:44PM -0800, Paul E. McKenney wrote:
> > rcu: Enable tick for nohz_full CPUs slow to provide expedited QS
> >
> > An expedited grace period can be stalled by a nohz_full CPU looping
> > in kernel context. This possibility is currently handled by some
> > carefully crafted checks in rcu_read_unlock_special() that enlist help
> > from ksoftirqd when permitted by the scheduler. However, it is exactly
> > these checks that require the scheduler avoid holding any of its rq or
> > pi locks across rcu_read_unlock() without also having held them across
> > the entire RCU read-side critical section.
>
> Just got a bit more free to look into the topic of RCU-related scheduler
> deadlocks I was previously looking into (And I will continue working on the
> WIP patch for detecting bad scenarios).
>
> > It would therefore be very nice if expedited grace periods could
> > handle nohz_full CPUs looping in kernel context without such checks.
>
> Could you clarify 'without such checks'? Are you referring to these checks in
> _unlock_special()?
> if (irqs_were_disabled && use_softirq &&
> (in_interrupt() ||
> (exp && !t->rcu_read_unlock_special.b.deferred_qs))) {
> // Using softirq, safe to awaken, and we get
> // no help from enabling irqs, unlike bh/preempt.
> raise_softirq_irqoff(RCU_SOFTIRQ);
>
> I am sorry but I did not fully understand this patch motivation but -- you
> are just saying you want EXP GPs to be handled quickly on nohz_full CPUs,
> right? Are we not turning the tick on already in rcu_nmi_enter_common() for
> handling slow EXP GPs as well? And if not, can we do it from there instead?

Currently, if the scheduler holds one of its rq or pi locks across
rcu_read_unlock(), it must do so across the entire RCU read-side critical
section. Although this odd requirement isn't a flaming disaster, it
would be good to lift if feasible. According to Lai's analysis (which
seems correct to me), the only remaining need for this requirement is the
raise_softirq_irqoff() you call out above. If expedited grace periods
can turn on the tick for holdout nohz_full CPUs, it seems to me that the
above check can be changed so as to avoid invoking raise_softirq_irqoff()
except in cases where it doesn't go into the scheduler. This could be
done by removing the portion of the "if" condition following the "||".

> > This commit therefore adds code to the expedited grace period's wait
> > and cleanup code that forces the scheduler-clock interrupt on for CPUs
> > that fail to quickly supply a quiescent state. "Quickly" is currently
> > a hard-coded single-jiffy delay.
> >
> > Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
> >
> > ---
> >
> > This is needed to avoid rcu_read_unlock_special() needing to enter
> > the scheduler for the benefit of expedited grace periods on nohz_full
> > CPUs, thus enabling more of Lai Jiangshan's patchset.
>
> Still need to go through Lai's patches :-(. Will be looking to make more time
> for the same.

It would indeed be good to get another set of eyes on this, as it is a
bit on the subtle side!

Thanx, Paul

> thanks!
>
> - Joel
>
>
> > diff --git a/include/linux/tick.h b/include/linux/tick.h
> > index 4ed788c..72a2a21 100644
> > --- a/include/linux/tick.h
> > +++ b/include/linux/tick.h
> > @@ -109,9 +109,10 @@ enum tick_dep_bits {
> > TICK_DEP_BIT_PERF_EVENTS = 1,
> > TICK_DEP_BIT_SCHED = 2,
> > TICK_DEP_BIT_CLOCK_UNSTABLE = 3,
> > - TICK_DEP_BIT_RCU = 4
> > + TICK_DEP_BIT_RCU = 4,
> > + TICK_DEP_BIT_RCU_EXP = 5
> > };
> > -#define TICK_DEP_BIT_MAX TICK_DEP_BIT_RCU
> > +#define TICK_DEP_BIT_MAX TICK_DEP_BIT_RCU_EXP
> >
> > #define TICK_DEP_MASK_NONE 0
> > #define TICK_DEP_MASK_POSIX_TIMER (1 << TICK_DEP_BIT_POSIX_TIMER)
> > @@ -119,6 +120,7 @@ enum tick_dep_bits {
> > #define TICK_DEP_MASK_SCHED (1 << TICK_DEP_BIT_SCHED)
> > #define TICK_DEP_MASK_CLOCK_UNSTABLE (1 << TICK_DEP_BIT_CLOCK_UNSTABLE)
> > #define TICK_DEP_MASK_RCU (1 << TICK_DEP_BIT_RCU)
> > +#define TICK_DEP_MASK_RCU_EXP (1 << TICK_DEP_BIT_RCU_EXP)
> >
> > #ifdef CONFIG_NO_HZ_COMMON
> > extern bool tick_nohz_enabled;
> > diff --git a/kernel/rcu/tree.h b/kernel/rcu/tree.h
> > index 634c1db..0c87e4c 100644
> > --- a/kernel/rcu/tree.h
> > +++ b/kernel/rcu/tree.h
> > @@ -181,6 +181,7 @@ struct rcu_data {
> > bool rcu_need_heavy_qs; /* GP old, so heavy quiescent state! */
> > bool rcu_urgent_qs; /* GP old need light quiescent state. */
> > bool rcu_forced_tick; /* Forced tick to provide QS. */
> > + bool rcu_forced_tick_exp; /* ... provide QS to expedited GP. */
> > #ifdef CONFIG_RCU_FAST_NO_HZ
> > unsigned long last_accelerate; /* Last jiffy CBs were accelerated. */
> > unsigned long last_advance_all; /* Last jiffy CBs were all advanced. */
> > diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
> > index 726ba20..6935a9e 100644
> > --- a/kernel/rcu/tree_exp.h
> > +++ b/kernel/rcu/tree_exp.h
> > @@ -230,7 +230,9 @@ static void __maybe_unused rcu_report_exp_rnp(struct rcu_node *rnp, bool wake)
> > static void rcu_report_exp_cpu_mult(struct rcu_node *rnp,
> > unsigned long mask, bool wake)
> > {
> > + int cpu;
> > unsigned long flags;
> > + struct rcu_data *rdp;
> >
> > raw_spin_lock_irqsave_rcu_node(rnp, flags);
> > if (!(rnp->expmask & mask)) {
> > @@ -238,6 +240,13 @@ static void rcu_report_exp_cpu_mult(struct rcu_node *rnp,
> > return;
> > }
> > WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
> > + for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
> > + rdp = per_cpu_ptr(&rcu_data, cpu);
> > + if (!IS_ENABLED(CONFIG_NO_HZ_FULL) || !rdp->rcu_forced_tick_exp)
> > + continue;
> > + rdp->rcu_forced_tick_exp = false;
> > + tick_dep_clear_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
> > + }
> > __rcu_report_exp_rnp(rnp, wake, flags); /* Releases rnp->lock. */
> > }
> >
> > @@ -450,6 +459,26 @@ static void sync_rcu_exp_select_cpus(void)
> > }
> >
> > /*
> > + * Wait for the expedited grace period to elapse, within time limit.
> > + * If the time limit is exceeded without the grace period elapsing,
> > + * return false, otherwise return true.
> > + */
> > +static bool synchronize_rcu_expedited_wait_once(long tlimit)
> > +{
> > + int t;
> > + struct rcu_node *rnp_root = rcu_get_root();
> > +
> > + t = swait_event_timeout_exclusive(rcu_state.expedited_wq,
> > + sync_rcu_exp_done_unlocked(rnp_root),
> > + tlimit);
> > + // Workqueues should not be signaled.
> > + if (t > 0 || sync_rcu_exp_done_unlocked(rnp_root))
> > + return true;
> > + WARN_ON(t < 0); /* workqueues should not be signaled. */
> > + return false;
> > +}
> > +
> > +/*
> > * Wait for the expedited grace period to elapse, issuing any needed
> > * RCU CPU stall warnings along the way.
> > */
> > @@ -460,22 +489,31 @@ static void synchronize_rcu_expedited_wait(void)
> > unsigned long jiffies_start;
> > unsigned long mask;
> > int ndetected;
> > + struct rcu_data *rdp;
> > struct rcu_node *rnp;
> > struct rcu_node *rnp_root = rcu_get_root();
> > - int ret;
> >
> > trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("startwait"));
> > jiffies_stall = rcu_jiffies_till_stall_check();
> > jiffies_start = jiffies;
> > + if (IS_ENABLED(CONFIG_NO_HZ_FULL)) {
> > + if (synchronize_rcu_expedited_wait_once(1))
> > + return;
> > + rcu_for_each_leaf_node(rnp) {
> > + for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
> > + rdp = per_cpu_ptr(&rcu_data, cpu);
> > + if (rdp->rcu_forced_tick_exp)
> > + continue;
> > + rdp->rcu_forced_tick_exp = true;
> > + tick_dep_set_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
> > + }
> > + }
> > + WARN_ON_ONCE(1);
> > + }
> >
> > for (;;) {
> > - ret = swait_event_timeout_exclusive(
> > - rcu_state.expedited_wq,
> > - sync_rcu_exp_done_unlocked(rnp_root),
> > - jiffies_stall);
> > - if (ret > 0 || sync_rcu_exp_done_unlocked(rnp_root))
> > + if (synchronize_rcu_expedited_wait_once(jiffies_stall))
> > return;
> > - WARN_ON(ret < 0); /* workqueues should not be signaled. */
> > if (rcu_cpu_stall_suppress)
> > continue;
> > panic_on_rcu_stall();

\
 
 \ /
  Last update: 2019-12-06 04:11    [W:0.038 / U:0.988 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site