On Fri, Aug 02, 2019 at 08:14:49AM -0700, Paul E. McKenney wrote: > Use of the rcu_data structure's segmented ->cblist for no-CBs CPUs > takes advantage of unrelated grace periods, thus reducing the memory > footprint in the face of floods of call_rcu() invocations. However, > the ->cblist field is a more-complex rcu_segcblist structure which must > be protected via locking. Even though there are only three entities > which can acquire this lock (the CPU invoking call_rcu(), the no-CBs > grace-period kthread, and the no-CBs callbacks kthread), the contention > on this lock is excessive under heavy stress. > > This commit therefore greatly reduces contention by provisioning > an rcu_cblist structure field named ->nocb_bypass within the > rcu_data structure. Each no-CBs CPU is permitted only a limited > number of enqueues onto the ->cblist per jiffy, controlled by a new > nocb_nobypass_lim_per_jiffy kernel boot parameter that defaults to > about 16 enqueues per millisecond (16 * 1000 / HZ). When that limit is > exceeded, the CPU instead enqueues onto the new ->nocb_bypass. Looks quite interesting. I am guessing the not-no-CB (regular) enqueues don't need to use the same technique because both enqueues / callback execution are happening on same CPU.. Still looking through patch but I understood the basic idea. Some nits below: [snip] > diff --git a/kernel/rcu/tree.h b/kernel/rcu/tree.h > index 2c3e9068671c..e4df86db8137 100644 > --- a/kernel/rcu/tree.h > +++ b/kernel/rcu/tree.h > @@ -200,18 +200,26 @@ struct rcu_data { > atomic_t nocb_lock_contended; /* Contention experienced. */ > int nocb_defer_wakeup; /* Defer wakeup of nocb_kthread. */ > struct timer_list nocb_timer; /* Enforce finite deferral. */ > + unsigned long nocb_gp_adv_time; /* Last call_rcu() CB adv (jiffies). */ > + > + /* The following fields are used by call_rcu, hence own cacheline. */ > + raw_spinlock_t nocb_bypass_lock ____cacheline_internodealigned_in_smp; > + struct rcu_cblist nocb_bypass; /* Lock-contention-bypass CB list. */ > + unsigned long nocb_bypass_first; /* Time (jiffies) of first enqueue. */ > + unsigned long nocb_nobypass_last; /* Last ->cblist enqueue (jiffies). */ > + int nocb_nobypass_count; /* # ->cblist enqueues at ^^^ time. */ Can these and below fields be ifdef'd out if !CONFIG_RCU_NOCB_CPU so as to keep the size of struct smaller for benefit of systems that don't use NOCB? > > /* The following fields are used by GP kthread, hence own cacheline. */ > raw_spinlock_t nocb_gp_lock ____cacheline_internodealigned_in_smp; > - bool nocb_gp_sleep; > - /* Is the nocb GP thread asleep? */ > + struct timer_list nocb_bypass_timer; /* Force nocb_bypass flush. */ > + bool nocb_gp_sleep; /* Is the nocb GP thread asleep? */ And these too, I think. > struct swait_queue_head nocb_gp_wq; /* For nocb kthreads to sleep on. */ > bool nocb_cb_sleep; /* Is the nocb CB thread asleep? */ > struct task_struct *nocb_cb_kthread; > struct rcu_data *nocb_next_cb_rdp; > /* Next rcu_data in wakeup chain. */ > > - /* The following fields are used by CB kthread, hence new cachline. */ > + /* The following fields are used by CB kthread, hence new cacheline. */ > struct rcu_data *nocb_gp_rdp ____cacheline_internodealigned_in_smp; > /* GP rdp takes GP-end wakeups. */ > #endif /* #ifdef CONFIG_RCU_NOCB_CPU */ [snip] > +static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j) > +{ > + rcu_lockdep_assert_cblist_protected(rdp); > + if (!rcu_segcblist_is_offloaded(&rdp->cblist) || > + !rcu_nocb_bypass_trylock(rdp)) > + return; > + WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j)); > +} > + > +/* > + * See whether it is appropriate to use the ->nocb_bypass list in order > + * to control contention on ->nocb_lock. A limited number of direct > + * enqueues are permitted into ->cblist per jiffy. If ->nocb_bypass > + * is non-empty, further callbacks must be placed into ->nocb_bypass, > + * otherwise rcu_barrier() breaks. Use rcu_nocb_flush_bypass() to switch > + * back to direct use of ->cblist. However, ->nocb_bypass should not be > + * used if ->cblist is empty, because otherwise callbacks can be stranded > + * on ->nocb_bypass because we cannot count on the current CPU ever again > + * invoking call_rcu(). The general rule is that if ->nocb_bypass is > + * non-empty, the corresponding no-CBs grace-period kthread must not be > + * in an indefinite sleep state. > + * > + * Finally, it is not permitted to use the bypass during early boot, > + * as doing so would confuse the auto-initialization code. Besides > + * which, there is no point in worrying about lock contention while > + * there is only one CPU in operation. > + */ > +static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp, > + bool *was_alldone, unsigned long flags) > +{ > + unsigned long c; > + unsigned long cur_gp_seq; > + unsigned long j = jiffies; > + long ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass); > + > + if (!rcu_segcblist_is_offloaded(&rdp->cblist)) { > + *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist); > + return false; /* Not offloaded, no bypassing. */ > + } > + lockdep_assert_irqs_disabled(); > + > + // Don't use ->nocb_bypass during early boot. Very minor nit: comment style should be /* */ thanks, - Joel [snip]