Hi Paul,
Some silly doubts.
On Mon, Sep 10, 2007 at 11:39:01AM -0700, Paul E. McKenney wrote:
> Work in progress, not for inclusion.
>
> RCU priority boosting is needed when running a workload that might include
> CPU-bound user tasks running at realtime priorities with preemptible RCU.
> In this situation, RCU priority boosting is needed to avoid OOM.
>
> Please note that because Classic RCU does not permit RCU read-side
> critical sections to be preempted, there is no need to boost the priority
> of Classic RCU readers. Boosting the priority of a running process
> does not make it run any faster, at least not on any hardware that I am
> aware of. ;-)
>
> Signed-off-by: Paul E. McKenney <[email protected]>
> ---
>
> include/linux/init_task.h | 13
> include/linux/rcupdate.h | 17 +
> include/linux/rcupreempt.h | 20 +
> include/linux/sched.h | 24 +
> init/main.c | 1
> kernel/Kconfig.preempt | 14 -
> kernel/fork.c | 6
> kernel/rcupreempt.c | 608 ++++++++++++++++++++++++++++++++++++++++++---
> kernel/rtmutex.c | 7
> kernel/sched.c | 5
> lib/Kconfig.debug | 34 ++
> 11 files changed, 703 insertions(+), 46 deletions(-)
>
> diff -urpNa -X dontdiff linux-2.6.22-E-hotplug/include/linux/init_task.h linux-2.6.22-F-boostrcu/include/linux/init_task.h
> --- linux-2.6.22-E-hotplug/include/linux/init_task.h 2007-07-08 16:32:17.000000000 -0700
> +++ linux-2.6.22-F-boostrcu/include/linux/init_task.h 2007-08-31 14:09:02.000000000 -0700
> @@ -87,6 +87,17 @@ extern struct nsproxy init_nsproxy;
> .signalfd_list = LIST_HEAD_INIT(sighand.signalfd_list), \
> }
>
> +#ifdef CONFIG_PREEMPT_RCU_BOOST
> +#define INIT_RCU_BOOST_PRIO .rcu_prio = MAX_PRIO,
> +#define INIT_PREEMPT_RCU_BOOST(tsk) \
> + .rcub_rbdp = NULL, \
> + .rcub_state = RCU_BOOST_IDLE, \
> + .rcub_entry = LIST_HEAD_INIT(tsk.rcub_entry),
> +#else /* #ifdef CONFIG_PREEMPT_RCU_BOOST */
> +#define INIT_RCU_BOOST_PRIO
> +#define INIT_PREEMPT_RCU_BOOST(tsk)
> +#endif /* #else #ifdef CONFIG_PREEMPT_RCU_BOOST */
> +
> extern struct group_info init_groups;
>
> #define INIT_STRUCT_PID { \
> @@ -125,6 +136,7 @@ extern struct group_info init_groups;
> .prio = MAX_PRIO-20, \
> .static_prio = MAX_PRIO-20, \
> .normal_prio = MAX_PRIO-20, \
> + INIT_RCU_BOOST_PRIO \
> .policy = SCHED_NORMAL, \
> .cpus_allowed = CPU_MASK_ALL, \
> .mm = NULL, \
> @@ -169,6 +181,7 @@ extern struct group_info init_groups;
> }, \
> INIT_TRACE_IRQFLAGS \
> INIT_LOCKDEP \
> + INIT_PREEMPT_RCU_BOOST(tsk) \
> }
>
>
> diff -urpNa -X dontdiff linux-2.6.22-E-hotplug/include/linux/rcupdate.h linux-2.6.22-F-boostrcu/include/linux/rcupdate.h
> --- linux-2.6.22-E-hotplug/include/linux/rcupdate.h 2007-08-24 11:03:22.000000000 -0700
> +++ linux-2.6.22-F-boostrcu/include/linux/rcupdate.h 2007-08-24 17:04:14.000000000 -0700
> @@ -252,5 +252,22 @@ static inline void rcu_qsctr_inc(int cpu
> per_cpu(rcu_data_passed_quiesc, cpu) = 1;
> }
>
> +#ifdef CONFIG_PREEMPT_RCU_BOOST
> +extern void init_rcu_boost_late(void);
> +extern void __rcu_preempt_boost(void);
> +#define rcu_preempt_boost() /* cpp to avoid #include hell. */ \
> + do { \
> + if (unlikely(current->rcu_read_lock_nesting > 0)) \
> + __rcu_preempt_boost(); \
> + } while (0)
> +#else /* #ifdef CONFIG_PREEMPT_RCU_BOOST */
> +static inline void init_rcu_boost_late(void)
> +{
> +}
> +static inline void rcu_preempt_boost(void)
> +{
> +}
> +#endif /* #else #ifdef CONFIG_PREEMPT_RCU_BOOST */
> +
> #endif /* __KERNEL__ */
> #endif /* __LINUX_RCUPDATE_H */
> diff -urpNa -X dontdiff linux-2.6.22-E-hotplug/include/linux/rcupreempt.h linux-2.6.22-F-boostrcu/include/linux/rcupreempt.h
> --- linux-2.6.22-E-hotplug/include/linux/rcupreempt.h 2007-08-24 11:20:32.000000000 -0700
> +++ linux-2.6.22-F-boostrcu/include/linux/rcupreempt.h 2007-08-24 11:24:59.000000000 -0700
> @@ -42,6 +42,26 @@
> #include <linux/cpumask.h>
> #include <linux/seqlock.h>
>
> +#ifdef CONFIG_PREEMPT_RCU_BOOST
> +/*
> + * Task state with respect to being RCU-boosted. This state is changed
> + * by the task itself in response to the following three events:
^^^
> + * 1. Preemption (or block on lock) while in RCU read-side critical section.
I am wondering, can a task block on a lock while in RCU read-side
critical section?
> + * 2. Outermost rcu_read_unlock() for blocked RCU read-side critical section.
> + *
Event #3. is missing?
The state change from RCU_BOOST_BLOCKED to RCU_BOOSTED is not done by
the task itself, but by the rcu_boost_task. No?
> + * The RCU-boost task also updates the state when boosting priority.
> + */
> +enum rcu_boost_state {
> + RCU_BOOST_IDLE = 0, /* Not yet blocked if in RCU read-side. */
> + RCU_BOOST_BLOCKED = 1, /* Blocked from RCU read-side. */
> + RCU_BOOSTED = 2, /* Boosting complete. */
> + RCU_BOOST_INVALID = 3, /* For bogus state sightings. */
> +};
> +
> +#define N_RCU_BOOST_STATE (RCU_BOOST_INVALID + 1)
> +
> +#endif /* #ifdef CONFIG_PREEMPT_RCU_BOOST */
> +
> #define call_rcu_bh(head, rcu) call_rcu(head, rcu)
> #define rcu_bh_qsctr_inc(cpu) do { } while (0)
> #define __rcu_read_lock_bh() { rcu_read_lock(); local_bh_disable(); }
> diff -urpNa -X dontdiff linux-2.6.22-E-hotplug/include/linux/sched.h linux-2.6.22-F-boostrcu/include/linux/sched.h
> --- linux-2.6.22-E-hotplug/include/linux/sched.h 2007-08-24 11:00:39.000000000 -0700
> +++ linux-2.6.22-F-boostrcu/include/linux/sched.h 2007-08-24 17:07:01.000000000 -0700
> @@ -546,6 +546,22 @@ struct signal_struct {
> #define is_rt_policy(p) ((p) != SCHED_NORMAL && (p) != SCHED_BATCH)
> #define has_rt_policy(p) unlikely(is_rt_policy((p)->policy))
>
> +#ifdef CONFIG_PREEMPT_RCU_BOOST
> +#define set_rcu_prio(p, prio) /* cpp to avoid #include hell */ \
> + do { \
> + (p)->rcu_prio = (prio); \
> + } while (0)
> +#define get_rcu_prio(p) (p)->rcu_prio /* cpp to avoid #include hell */
> +#else /* #ifdef CONFIG_PREEMPT_RCU_BOOST */
> +static inline void set_rcu_prio(struct task_struct *p, int prio)
> +{
> +}
> +static inline int get_rcu_prio(struct task_struct *p)
> +{
> + return MAX_PRIO;
> +}
> +#endif /* #else #ifdef CONFIG_PREEMPT_RCU_BOOST */
> +
> /*
> * Some day this will be a full-fledged user tracking system..
> */
> @@ -834,6 +850,9 @@ struct task_struct {
> #endif
> int load_weight; /* for niceness load balancing purposes */
> int prio, static_prio, normal_prio;
> +#ifdef CONFIG_PREEMPT_RCU_BOOST
> + int rcu_prio;
> +#endif /* #ifdef CONFIG_PREEMPT_RCU_BOOST */
> struct list_head run_list;
> struct prio_array *array;
>
> @@ -858,6 +877,11 @@ struct task_struct {
> #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
> struct sched_info sched_info;
> #endif
> +#ifdef CONFIG_PREEMPT_RCU_BOOST
> + struct rcu_boost_dat *rcub_rbdp;
> + enum rcu_boost_state rcub_state;
> + struct list_head rcub_entry;
> +#endif /* #ifdef CONFIG_PREEMPT_RCU_BOOST */
>
> struct list_head tasks;
> /*
> diff -urpNa -X dontdiff linux-2.6.22-E-hotplug/init/main.c linux-2.6.22-F-boostrcu/init/main.c
> --- linux-2.6.22-E-hotplug/init/main.c 2007-07-08 16:32:17.000000000 -0700
> +++ linux-2.6.22-F-boostrcu/init/main.c 2007-08-24 11:24:59.000000000 -0700
> @@ -722,6 +722,7 @@ static void __init do_basic_setup(void)
> driver_init();
> init_irq_proc();
> do_initcalls();
> + init_rcu_boost_late();
> }
>
> static void __init do_pre_smp_initcalls(void)
> diff -urpNa -X dontdiff linux-2.6.22-E-hotplug/kernel/fork.c linux-2.6.22-F-boostrcu/kernel/fork.c
> --- linux-2.6.22-E-hotplug/kernel/fork.c 2007-08-24 11:00:39.000000000 -0700
> +++ linux-2.6.22-F-boostrcu/kernel/fork.c 2007-08-24 11:24:59.000000000 -0700
> @@ -1036,6 +1036,12 @@ static struct task_struct *copy_process(
> p->rcu_read_lock_nesting = 0;
> p->rcu_flipctr_idx = 0;
> #endif /* #ifdef CONFIG_PREEMPT_RCU */
> +#ifdef CONFIG_PREEMPT_RCU_BOOST
> + p->rcu_prio = MAX_PRIO;
> + p->rcub_rbdp = NULL;
> + p->rcub_state = RCU_BOOST_IDLE;
> + INIT_LIST_HEAD(&p->rcub_entry);
> +#endif /* #ifdef CONFIG_PREEMPT_RCU_BOOST */
> p->vfork_done = NULL;
> spin_lock_init(&p->alloc_lock);
>
> diff -urpNa -X dontdiff linux-2.6.22-E-hotplug/kernel/Kconfig.preempt linux-2.6.22-F-boostrcu/kernel/Kconfig.preempt
> --- linux-2.6.22-E-hotplug/kernel/Kconfig.preempt 2007-08-24 11:00:39.000000000 -0700
> +++ linux-2.6.22-F-boostrcu/kernel/Kconfig.preempt 2007-08-24 11:24:59.000000000 -0700
> @@ -91,13 +91,13 @@ config PREEMPT_RCU
>
> endchoice
>
> -config RCU_TRACE
> - bool "Enable tracing for RCU - currently stats in debugfs"
> - select DEBUG_FS
> - default y
> +config PREEMPT_RCU_BOOST
> + bool "Enable priority boosting of RCU read-side critical sections"
> + depends on PREEMPT_RCU
> + default n
> help
> - This option provides tracing in RCU which presents stats
> - in debugfs for debugging RCU implementation.
> + This option permits priority boosting of RCU read-side critical
> + sections that have been preempted in order to prevent indefinite
> + delay of grace periods in face of runaway non-realtime processes.
>
> - Say Y here if you want to enable RCU tracing
> Say N if you are unsure.
> diff -urpNa -X dontdiff linux-2.6.22-E-hotplug/kernel/rcupreempt.c linux-2.6.22-F-boostrcu/kernel/rcupreempt.c
> --- linux-2.6.22-E-hotplug/kernel/rcupreempt.c 2007-08-24 11:20:32.000000000 -0700
> +++ linux-2.6.22-F-boostrcu/kernel/rcupreempt.c 2007-08-31 14:06:49.000000000 -0700
> @@ -51,6 +51,15 @@
> #include <linux/byteorder/swabb.h>
> #include <linux/cpumask.h>
> #include <linux/rcupreempt_trace.h>
> +#include <linux/kthread.h>
> +
> +/*
> + * Macro that prevents the compiler from reordering accesses, but does
> + * absolutely -nothing- to prevent CPUs from reordering. This is used
> + * only to mediate communication between mainline code and hardware
> + * interrupt and NMI handlers.
> + */
> +#define ORDERED_WRT_IRQ(x) (*(volatile typeof(x) *)&(x))
>
> /*
> * PREEMPT_RCU data structures.
> @@ -82,6 +91,531 @@ static struct rcu_ctrlblk rcu_ctrlblk =
> };
> static DEFINE_PER_CPU(int [2], rcu_flipctr) = { 0, 0 };
>
> +#ifndef CONFIG_PREEMPT_RCU_BOOST
> +static inline void init_rcu_boost_early(void) { }
> +static inline void rcu_read_unlock_unboost(void) { }
> +#else /* #ifndef CONFIG_PREEMPT_RCU_BOOST */
> +
> +/* Defines possible event indices for ->rbs_stats[] (first index). */
> +
> +#define RCU_BOOST_DAT_BLOCK 0
> +#define RCU_BOOST_DAT_BOOST 1
> +#define RCU_BOOST_DAT_UNLOCK 2
> +#define N_RCU_BOOST_DAT_EVENTS 3
> +
> +/* RCU-boost per-CPU array element. */
> +
> +struct rcu_boost_dat {
> + spinlock_t rbs_lock; /* Protects state/CPU slice of structures. */
> + struct list_head rbs_toboost;
> + struct list_head rbs_boosted;
> + unsigned long rbs_blocked;
> + unsigned long rbs_boost_attempt;
> + unsigned long rbs_boost;
> + unsigned long rbs_unlock;
> + unsigned long rbs_unboosted;
> +#ifdef CONFIG_PREEMPT_RCU_BOOST_STATS
> + unsigned long rbs_stats[N_RCU_BOOST_DAT_EVENTS][N_RCU_BOOST_STATE];
> +#endif /* #ifdef CONFIG_PREEMPT_RCU_BOOST_STATS */
> +};
> +#define RCU_BOOST_ELEMENTS 4
> +
> +static int rcu_boost_idx = -1; /* invalid value for early RCU use. */
> +static DEFINE_PER_CPU(struct rcu_boost_dat, rcu_boost_dat[RCU_BOOST_ELEMENTS]);
> +static struct task_struct *rcu_boost_task;
> +
> +#ifdef CONFIG_PREEMPT_RCU_BOOST_STATS
> +
> +/*
> + * Function to increment indicated ->rbs_stats[] element.
> + */
> +static inline void rcu_boost_dat_stat(struct rcu_boost_dat *rbdp,
> + int event,
> + enum rcu_boost_state oldstate)
> +{
> + if (oldstate >= RCU_BOOST_IDLE && oldstate <= RCU_BOOSTED) {
> + rbdp->rbs_stats[event][oldstate]++;
> + } else {
> + rbdp->rbs_stats[event][RCU_BOOST_INVALID]++;
> + }
> +}
> +
> +static inline void rcu_boost_dat_stat_block(struct rcu_boost_dat *rbdp,
> + enum rcu_boost_state oldstate)
> +{
> + rcu_boost_dat_stat(rbdp, RCU_BOOST_DAT_BLOCK, oldstate);
> +}
> +
> +static inline void rcu_boost_dat_stat_boost(struct rcu_boost_dat *rbdp,
> + enum rcu_boost_state oldstate)
> +{
> + rcu_boost_dat_stat(rbdp, RCU_BOOST_DAT_BOOST, oldstate);
> +}
> +
> +static inline void rcu_boost_dat_stat_unlock(struct rcu_boost_dat *rbdp,
> + enum rcu_boost_state oldstate)
> +{
> + rcu_boost_dat_stat(rbdp, RCU_BOOST_DAT_UNLOCK, oldstate);
> +}
> +
> +/*
> + * Prefix for kprint() strings for periodic statistics messages.
> + */
> +static char *rcu_boost_state_event[] = {
> + "block: ",
> + "boost: ",
> + "unlock: ",
> +};
> +
> +/*
> + * Indicators for numbers in kprint() strings. "!" indicates a state-event
> + * pair that should not happen, while "?" indicates a state that should
> + * not happen.
> + */
> +static char *rcu_boost_state_error[] = {
> + /*ibBe*/
> + " ?", /* block */
> + "! ?", /* boost */
> + "? ?", /* unlock */
> +};
> +
> +/*
> + * Print out RCU booster task statistics at the specified interval.
> + */
> +static void rcu_boost_dat_stat_print(void)
> +{
> + /* Three decimal digits per byte plus spacing per number and line. */
> + char buf[N_RCU_BOOST_STATE * (sizeof(long) * 3 + 2) + 2];
> + int cpu;
> + int event;
> + int i;
> + static time_t lastprint; /* static implies 0 initial value. */
> + struct rcu_boost_dat *rbdp;
> + int state;
> + struct rcu_boost_dat sum;
> +
> + /* Wait a graceful interval between printk spamming. */
> + /* Note: time_after() dislikes time_t. */
> +
> + if (xtime.tv_sec - lastprint < CONFIG_PREEMPT_RCU_BOOST_STATS_INTERVAL)
> + return;
> +
> + /* Sum up the state/event-independent counters. */
> +
> + sum.rbs_blocked = 0;
> + sum.rbs_boost_attempt = 0;
> + sum.rbs_boost = 0;
> + sum.rbs_unlock = 0;
> + sum.rbs_unboosted = 0;
> + for_each_possible_cpu(cpu)
> + for (i = 0; i < RCU_BOOST_ELEMENTS; i++) {
> + rbdp = per_cpu(rcu_boost_dat, cpu);
> + sum.rbs_blocked += rbdp[i].rbs_blocked;
> + sum.rbs_boost_attempt += rbdp[i].rbs_boost_attempt;
> + sum.rbs_boost += rbdp[i].rbs_boost;
> + sum.rbs_unlock += rbdp[i].rbs_unlock;
> + sum.rbs_unboosted += rbdp[i].rbs_unboosted;
> + }
> +
> + /* Sum up the state/event-dependent counters. */
> +
> + for (event = 0; event < N_RCU_BOOST_DAT_EVENTS; event++)
> + for (state = 0; state < N_RCU_BOOST_STATE; state++) {
> + sum.rbs_stats[event][state] = 0;
> + for_each_possible_cpu(cpu) {
> + for (i = 0; i < RCU_BOOST_ELEMENTS; i++)
> + sum.rbs_stats[event][state]
> + += per_cpu(rcu_boost_dat,
> + cpu)[i].rbs_stats[event][state];
> + }
> + }
> +
> + /* Print them out! */
> +
> + printk(KERN_INFO
> + "rcu_boost_dat: idx=%d "
> + "b=%lu ul=%lu ub=%lu boost: a=%lu b=%lu\n",
> + rcu_boost_idx,
> + sum.rbs_blocked, sum.rbs_unlock, sum.rbs_unboosted,
> + sum.rbs_boost_attempt, sum.rbs_boost);
> + for (event = 0; event < N_RCU_BOOST_DAT_EVENTS; event++) {
> + i = 0;
> + for (state = 0; state < N_RCU_BOOST_STATE; state++)
> + i += sprintf(&buf[i], " %ld%c",
> + sum.rbs_stats[event][state],
> + rcu_boost_state_error[event][state]);
> + printk(KERN_INFO "rcu_boost_dat %s %s\n",
> + rcu_boost_state_event[event], buf);
> + }
> +
> + /* Go away and don't come back for awhile. */
> +
> + lastprint = xtime.tv_sec;
> +}
> +
> +#else /* #ifdef CONFIG_PREEMPT_RCU_BOOST_STATS */
> +
> +static inline void rcu_boost_dat_stat_block(struct rcu_boost_dat *rbdp,
> + enum rcu_boost_state oldstate)
> +{
> +}
> +static inline void rcu_boost_dat_stat_boost(struct rcu_boost_dat *rbdp,
> + enum rcu_boost_state oldstate)
> +{
> +}
> +static inline void rcu_boost_dat_stat_unlock(struct rcu_boost_dat *rbdp,
> + enum rcu_boost_state oldstate)
> +{
> +}
> +static void rcu_boost_dat_stat_print(void)
> +{
> +}
> +
> +#endif /* #else #ifdef CONFIG_PREEMPT_RCU_BOOST_STATS */
> +
> +/*
> + * Initialize RCU-boost state. This happens early in the boot process,
> + * when the scheduler does not yet exist. So don't try to use it.
> + */
> +static void init_rcu_boost_early(void)
> +{
> + struct rcu_boost_dat *rbdp;
> + int cpu;
> + int i;
> +
> + for_each_possible_cpu(cpu) {
> + rbdp = per_cpu(rcu_boost_dat, cpu);
> + for (i = 0; i < RCU_BOOST_ELEMENTS; i++) {
> + spin_lock_init(&rbdp[i].rbs_lock);
> + INIT_LIST_HEAD(&rbdp[i].rbs_toboost);
> + INIT_LIST_HEAD(&rbdp[i].rbs_boosted);
> + rbdp[i].rbs_blocked = 0;
> + rbdp[i].rbs_boost_attempt = 0;
> + rbdp[i].rbs_boost = 0;
> + rbdp[i].rbs_unlock = 0;
> + rbdp[i].rbs_unboosted = 0;
> +#ifdef CONFIG_PREEMPT_RCU_BOOST_STATS
> + {
> + int j, k;
> +
> + for (j = 0; j < N_RCU_BOOST_DAT_EVENTS; j++)
> + for (k = 0; k < N_RCU_BOOST_STATE; k++)
> + rbdp[i].rbs_stats[j][k] = 0;
> + }
> +#endif /* #ifdef CONFIG_PREEMPT_RCU_BOOST_STATS */
> + }
> + smp_wmb(); /* Make sure readers see above initialization. */
> + rcu_boost_idx = 0; /* Allow readers to access data. */
> + }
> +}
> +
> +/*
> + * Return the list on which the calling task should add itself, or
> + * NULL if too early during initialization.
> + */
> +static inline struct rcu_boost_dat *rcu_rbd_new(void)
> +{
> + int cpu = smp_processor_id(); /* locks used, so preemption OK. */
> + int idx = ORDERED_WRT_IRQ(rcu_boost_idx);
> +
> + if (unlikely(idx < 0))
> + return NULL;
> + return &per_cpu(rcu_boost_dat, cpu)[idx];
> +}
> +
> +/*
> + * Return the list from which to boost target tasks.
> + * May only be invoked by the booster task, so guaranteed to
> + * already be initialized. Use rcu_boost_dat element least recently
> + * the destination for task blocking in RCU read-side critical sections.
> + */
> +static inline struct rcu_boost_dat *rcu_rbd_boosting(int cpu)
> +{
> + int idx = (rcu_boost_idx + 1) & (RCU_BOOST_ELEMENTS - 1);
> +
> + return &per_cpu(rcu_boost_dat, cpu)[idx];
> +}
> +
> +#define PREEMPT_RCU_BOOSTER_PRIO 49 /* Match curr_irq_prio manually. */
> + /* Administrators can always adjust */
> + /* via the /proc interface. */
> +
> +/*
> + * Boost the specified task from an RCU viewpoint.
> + * Boost the target task to a priority just a bit less-favored than
> + * that of the RCU-boost task, but boost to a realtime priority even
> + * if the RCU-boost task is running at a non-realtime priority.
> + * We check the priority of the RCU-boost task each time we boost
> + * in case the sysadm manually changes the priority.
> + */
> +static void rcu_boost_prio(struct task_struct *taskp)
> +{
> + unsigned long flags;
> + int rcuprio;
> +
> + spin_lock_irqsave(¤t->pi_lock, flags);
> + rcuprio = rt_mutex_getprio(current) + 1;
> + if (rcuprio >= MAX_USER_RT_PRIO)
> + rcuprio = MAX_USER_RT_PRIO - 1;
> + spin_unlock_irqrestore(¤t->pi_lock, flags);
> + spin_lock_irqsave(&taskp->pi_lock, flags);
> + if (taskp->rcu_prio != rcuprio) {
> + taskp->rcu_prio = rcuprio;
> + if (taskp->rcu_prio < taskp->prio)
> + rt_mutex_setprio(taskp, taskp->rcu_prio);
> + }
> + spin_unlock_irqrestore(&taskp->pi_lock, flags);
> +}
> +
> +/*
> + * Unboost the specified task from an RCU viewpoint.
> + */
> +static void rcu_unboost_prio(struct task_struct *taskp)
> +{
> + int nprio;
> + unsigned long flags;
> +
> + spin_lock_irqsave(&taskp->pi_lock, flags);
> + taskp->rcu_prio = MAX_PRIO;
> + nprio = rt_mutex_getprio(taskp);
> + if (nprio > taskp->prio)
> + rt_mutex_setprio(taskp, nprio);
> + spin_unlock_irqrestore(&taskp->pi_lock, flags);
> +}
> +
> +/*
> + * Boost all of the RCU-reader tasks on the specified list.
> + */
> +static void rcu_boost_one_reader_list(struct rcu_boost_dat *rbdp)
> +{
> + LIST_HEAD(list);
> + unsigned long flags;
> + struct task_struct *taskp;
> +
> + /*
> + * Splice both lists onto a local list. We will still
> + * need to hold the lock when manipulating the local list
> + * because tasks can remove themselves at any time.
> + * The reason for splicing the rbs_boosted list is that
> + * our priority may have changed, so reboosting may be
> + * required.
> + */
> +
> + spin_lock_irqsave(&rbdp->rbs_lock, flags);
> + list_splice_init(&rbdp->rbs_toboost, &list);
> + list_splice_init(&rbdp->rbs_boosted, &list);
> + while (!list_empty(&list)) {
> +
> + /*
> + * Pause for a bit before boosting each task.
> + * @@@FIXME: reduce/eliminate pausing in case of OOM.
> + */
> +
> + spin_unlock_irqrestore(&rbdp->rbs_lock, flags);
> + schedule_timeout_uninterruptible(1);
> + spin_lock_irqsave(&rbdp->rbs_lock, flags);
> +
> + /*
> + * All tasks might have removed themselves while
> + * we were waiting. Recheck list emptiness.
> + */
> +
> + if (list_empty(&list))
> + break;
> +
> + /* Remove first task in local list, count the attempt. */
> +
> + taskp = list_entry(list.next, typeof(*taskp), rcub_entry);
> + list_del_init(&taskp->rcub_entry);
> + rbdp->rbs_boost_attempt++;
> +
> + /* Ignore tasks in unexpected states. */
> +
> + if (taskp->rcub_state == RCU_BOOST_IDLE) {
> + list_add_tail(&taskp->rcub_entry, &rbdp->rbs_toboost);
> + rcu_boost_dat_stat_boost(rbdp, taskp->rcub_state);
> + continue;
> + }
> +
> + /* Boost the task's priority. */
> +
> + rcu_boost_prio(taskp);
> + rbdp->rbs_boost++;
> + rcu_boost_dat_stat_boost(rbdp, taskp->rcub_state);
> + taskp->rcub_state = RCU_BOOSTED;
> + list_add_tail(&taskp->rcub_entry, &rbdp->rbs_boosted);
> + }
> + spin_unlock_irqrestore(&rbdp->rbs_lock, flags);
> +}
> +
> +/*
> + * Priority-boost tasks stuck in RCU read-side critical sections as
> + * needed (presumably rarely).
> + */
> +static int rcu_booster(void *arg)
> +{
> + int cpu;
> + struct sched_param sp = { .sched_priority = PREEMPT_RCU_BOOSTER_PRIO, };
> +
> + sched_setscheduler(current, SCHED_RR, &sp);
> + current->flags |= PF_NOFREEZE;
> +
> + do {
> +
> + /* Advance the lists of tasks. */
> +
> + rcu_boost_idx = (rcu_boost_idx + 1) % RCU_BOOST_ELEMENTS;
> + for_each_possible_cpu(cpu) {
> +
> + /*
> + * Boost all sufficiently aged readers.
> + * Readers must first be preempted or block
> + * on a mutex in an RCU read-side critical section,
> + * then remain in that critical section for
> + * RCU_BOOST_ELEMENTS-1 time intervals.
> + * So most of the time we should end up doing
> + * nothing.
> + */
> +
> + rcu_boost_one_reader_list(rcu_rbd_boosting(cpu));
> +
> + /*
> + * Large SMP systems may need to sleep sometimes
> + * in this loop. Or have multiple RCU-boost tasks.
> + */
> + }
> +
> + /*
> + * Sleep to allow any unstalled RCU read-side critical
> + * sections to age out of the list. @@@ FIXME: reduce,
> + * adjust, or eliminate in case of OOM.
> + */
> +
> + schedule_timeout_uninterruptible(HZ);
> +
> + /* Print stats if enough time has passed. */
> +
> + rcu_boost_dat_stat_print();
> +
> + } while (!kthread_should_stop());
> +
> + return 0;
> +}
> +
> +/*
> + * Perform the portions of RCU-boost initialization that require the
> + * scheduler to be up and running.
> + */
> +void init_rcu_boost_late(void)
> +{
> +
> + /* Spawn RCU-boost task. */
> +
> + printk(KERN_INFO "Starting RCU priority booster\n");
> + rcu_boost_task = kthread_run(rcu_booster, NULL, "RCU Prio Booster");
> + if (IS_ERR(rcu_boost_task))
> + panic("Unable to create RCU Priority Booster, errno %ld\n",
> + -PTR_ERR(rcu_boost_task));
> +}
> +
> +/*
> + * Update task's RCU-boost state to reflect blocking in RCU read-side
> + * critical section, so that the RCU-boost task can find it in case it
> + * later needs its priority boosted.
> + */
> +void __rcu_preempt_boost(void)
> +{
> + struct rcu_boost_dat *rbdp;
> + unsigned long flags;
> +
> + /* Identify list to place task on for possible later boosting. */
> +
> + local_irq_save(flags);
> + rbdp = rcu_rbd_new();
> + if (rbdp == NULL) {
> + local_irq_restore(flags);
> + printk(KERN_INFO
> + "Preempted RCU read-side critical section too early.\n");
> + return;
> + }
> + spin_lock(&rbdp->rbs_lock);
> + rbdp->rbs_blocked++;
> +
> + /*
> + * Update state. We hold the lock and aren't yet on the list,
> + * so the booster cannot mess with us yet.
> + */
> +
> + rcu_boost_dat_stat_block(rbdp, current->rcub_state);
> + if (current->rcub_state != RCU_BOOST_IDLE) {
> +
> + /*
> + * We have been here before, so just update stats.
> + * It may seem strange to do all this work just to
> + * accumulate statistics, but this is such a
> + * low-probability code path that we shouldn't care.
> + * If it becomes a problem, it can be fixed.
> + */
> +
> + spin_unlock_irqrestore(&rbdp->rbs_lock, flags);
> + return;
> + }
> + current->rcub_state = RCU_BOOST_BLOCKED;
> +
> + /* Now add ourselves to the list so that the booster can find us. */
> +
> + list_add_tail(¤t->rcub_entry, &rbdp->rbs_toboost);
> + current->rcub_rbdp = rbdp;
> + spin_unlock_irqrestore(&rbdp->rbs_lock, flags);
> +}
> +
> +/*
> + * Do the list-removal and priority-unboosting "heavy lifting" when
> + * required.
> + */
> +static void __rcu_read_unlock_unboost(void)
> +{
> + unsigned long flags;
> + struct rcu_boost_dat *rbdp;
> +
> + /* Identify the list structure and acquire the corresponding lock. */
> +
> + rbdp = current->rcub_rbdp;
> + spin_lock_irqsave(&rbdp->rbs_lock, flags);
> +
> + /* Remove task from the list it was on. */
> +
> + list_del_init(¤t->rcub_entry);
> + rbdp->rbs_unlock++;
> + current->rcub_rbdp = NULL;
> +
> + /* Record stats, unboost if needed, and update state. */
> +
> + rcu_boost_dat_stat_unlock(rbdp, current->rcub_state);
> + if (current->rcub_state == RCU_BOOSTED) {
> + rcu_unboost_prio(current);
> + rbdp->rbs_unboosted++;
> + }
> + current->rcub_state = RCU_BOOST_IDLE;
> + spin_unlock_irqrestore(&rbdp->rbs_lock, flags);
> +}
> +
> +/*
> + * Do any state changes and unboosting needed for rcu_read_unlock().
> + * Pass any complex work on to __rcu_read_unlock_unboost().
> + * The vast majority of the time, no work will be needed, as preemption
> + * and blocking within RCU read-side critical sections is comparatively
> + * rare.
> + */
> +static inline void rcu_read_unlock_unboost(void)
> +{
> +
> + if (unlikely(current->rcub_state != RCU_BOOST_IDLE))
> + __rcu_read_unlock_unboost();
> +}
> +
> +#endif /* #else #ifndef CONFIG_PREEMPT_RCU_BOOST */
> +
> /*
> * States for rcu_try_flip() and friends.
> */
> @@ -128,14 +662,6 @@ static DEFINE_PER_CPU(enum rcu_mb_flag_v
> static cpumask_t rcu_cpu_online_map = CPU_MASK_NONE;
>
> /*
> - * Macro that prevents the compiler from reordering accesses, but does
> - * absolutely -nothing- to prevent CPUs from reordering. This is used
> - * only to mediate communication between mainline code and hardware
> - * interrupt and NMI handlers.
> - */
> -#define ORDERED_WRT_IRQ(x) (*(volatile typeof(x) *)&(x))
> -
> -/*
> * RCU_DATA_ME: find the current CPU's rcu_data structure.
> * RCU_DATA_CPU: find the specified CPU's rcu_data structure.
> */
> @@ -194,7 +720,7 @@ void __rcu_read_lock(void)
> me->rcu_read_lock_nesting = nesting + 1;
>
> } else {
> - unsigned long oldirq;
> + unsigned long flags;
>
> /*
> * Disable local interrupts to prevent the grace-period
> @@ -203,7 +729,7 @@ void __rcu_read_lock(void)
> * contain rcu_read_lock().
> */
>
> - local_irq_save(oldirq);
> + local_irq_save(flags);
>
> /*
> * Outermost nesting of rcu_read_lock(), so increment
> @@ -233,7 +759,7 @@ void __rcu_read_lock(void)
> */
>
> ORDERED_WRT_IRQ(me->rcu_flipctr_idx) = idx;
> - local_irq_restore(oldirq);
> + local_irq_restore(flags);
> }
> }
> EXPORT_SYMBOL_GPL(__rcu_read_lock);
> @@ -255,7 +781,7 @@ void __rcu_read_unlock(void)
> me->rcu_read_lock_nesting = nesting - 1;
>
> } else {
> - unsigned long oldirq;
> + unsigned long flags;
>
> /*
> * Disable local interrupts to prevent the grace-period
> @@ -264,7 +790,7 @@ void __rcu_read_unlock(void)
> * contain rcu_read_lock() and rcu_read_unlock().
> */
>
> - local_irq_save(oldirq);
> + local_irq_save(flags);
>
> /*
> * Outermost nesting of rcu_read_unlock(), so we must
> @@ -305,7 +831,10 @@ void __rcu_read_unlock(void)
> */
>
> ORDERED_WRT_IRQ(__get_cpu_var(rcu_flipctr)[idx])--;
> - local_irq_restore(oldirq);
> +
> + rcu_read_unlock_unboost();
> +
> + local_irq_restore(flags);
> }
> }
> EXPORT_SYMBOL_GPL(__rcu_read_unlock);
> @@ -504,10 +1033,10 @@ rcu_try_flip_waitmb(void)
> */
> static void rcu_try_flip(void)
> {
> - unsigned long oldirq;
> + unsigned long flags;
>
> RCU_TRACE_ME(rcupreempt_trace_try_flip_1);
> - if (unlikely(!spin_trylock_irqsave(&rcu_ctrlblk.fliplock, oldirq))) {
> + if (unlikely(!spin_trylock_irqsave(&rcu_ctrlblk.fliplock, flags))) {
> RCU_TRACE_ME(rcupreempt_trace_try_flip_e1);
> return;
> }
> @@ -534,7 +1063,7 @@ static void rcu_try_flip(void)
> if (rcu_try_flip_waitmb())
> rcu_try_flip_state = rcu_try_flip_idle_state;
> }
> - spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, oldirq);
> + spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
> }
>
> /*
> @@ -553,16 +1082,16 @@ static void rcu_check_mb(int cpu)
>
> void rcu_check_callbacks_rt(int cpu, int user)
> {
> - unsigned long oldirq;
> + unsigned long flags;
> struct rcu_data *rdp = RCU_DATA_CPU(cpu);
>
> rcu_check_mb(cpu);
> if (rcu_ctrlblk.completed == rdp->completed)
> rcu_try_flip();
> - spin_lock_irqsave(&rdp->lock, oldirq);
> + spin_lock_irqsave(&rdp->lock, flags);
> RCU_TRACE_RDP(rcupreempt_trace_check_callbacks, rdp);
> __rcu_advance_callbacks(rdp);
> - spin_unlock_irqrestore(&rdp->lock, oldirq);
> + spin_unlock_irqrestore(&rdp->lock, flags);
> }
>
> /*
> @@ -571,18 +1100,19 @@ void rcu_check_callbacks_rt(int cpu, int
> */
> void rcu_advance_callbacks_rt(int cpu, int user)
> {
> - unsigned long oldirq;
> + unsigned long flags;
> struct rcu_data *rdp = RCU_DATA_CPU(cpu);
>
> if (rcu_ctrlblk.completed == rdp->completed) {
> rcu_try_flip();
> if (rcu_ctrlblk.completed == rdp->completed)
> return;
> + rcu_read_unlock_unboost();
> }
> - spin_lock_irqsave(&rdp->lock, oldirq);
> + spin_lock_irqsave(&rdp->lock, flags);
> RCU_TRACE_RDP(rcupreempt_trace_check_callbacks, rdp);
> __rcu_advance_callbacks(rdp);
> - spin_unlock_irqrestore(&rdp->lock, oldirq);
> + spin_unlock_irqrestore(&rdp->lock, flags);
> }
>
> #ifdef CONFIG_HOTPLUG_CPU
> @@ -601,24 +1131,24 @@ void rcu_offline_cpu_rt(int cpu)
> {
> int i;
> struct rcu_head *list = NULL;
> - unsigned long oldirq;
> + unsigned long flags;
> struct rcu_data *rdp = RCU_DATA_CPU(cpu);
> struct rcu_head **tail = &list;
>
> /* Remove all callbacks from the newly dead CPU, retaining order. */
>
> - spin_lock_irqsave(&rdp->lock, oldirq);
> + spin_lock_irqsave(&rdp->lock, flags);
> rcu_offline_cpu_rt_enqueue(rdp->donelist, rdp->donetail, list, tail);
> for (i = GP_STAGES - 1; i >= 0; i--)
> rcu_offline_cpu_rt_enqueue(rdp->waitlist[i], rdp->waittail[i],
> list, tail);
> rcu_offline_cpu_rt_enqueue(rdp->nextlist, rdp->nexttail, list, tail);
> - spin_unlock_irqrestore(&rdp->lock, oldirq);
> + spin_unlock_irqrestore(&rdp->lock, flags);
> rdp->waitlistcount = 0;
>
> /* Disengage the newly dead CPU from grace-period computation. */
>
> - spin_lock_irqsave(&rcu_ctrlblk.fliplock, oldirq);
> + spin_lock_irqsave(&rcu_ctrlblk.fliplock, flags);
> rcu_check_mb(cpu);
> if (per_cpu(rcu_flip_flag, cpu) == rcu_flipped) {
> smp_mb(); /* Subsequent counter accesses must see new value */
> @@ -627,7 +1157,7 @@ void rcu_offline_cpu_rt(int cpu)
> /* seen -after- acknowledgement. */
> }
> cpu_clear(cpu, rcu_cpu_online_map);
> - spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, oldirq);
> + spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
>
> /*
> * Place the removed callbacks on the current CPU's queue.
> @@ -640,20 +1170,20 @@ void rcu_offline_cpu_rt(int cpu)
> */
>
> rdp = RCU_DATA_ME();
> - spin_lock_irqsave(&rdp->lock, oldirq);
> + spin_lock_irqsave(&rdp->lock, flags);
> *rdp->nexttail = list;
> if (list)
> rdp->nexttail = tail;
> - spin_unlock_irqrestore(&rdp->lock, oldirq);
> + spin_unlock_irqrestore(&rdp->lock, flags);
> }
>
> void __devinit rcu_online_cpu_rt(int cpu)
> {
> - unsigned long oldirq;
> + unsigned long flags;
>
> - spin_lock_irqsave(&rcu_ctrlblk.fliplock, oldirq);
> + spin_lock_irqsave(&rcu_ctrlblk.fliplock, flags);
> cpu_set(cpu, rcu_cpu_online_map);
> - spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, oldirq);
> + spin_unlock_irqrestore(&rcu_ctrlblk.fliplock, flags);
> }
>
> #else /* #ifdef CONFIG_HOTPLUG_CPU */
> @@ -695,12 +1225,12 @@ void rcu_process_callbacks_rt(struct sof
> void fastcall call_rcu_preempt(struct rcu_head *head,
> void (*func)(struct rcu_head *rcu))
> {
> - unsigned long oldirq;
> + unsigned long flags;
> struct rcu_data *rdp;
>
> head->func = func;
> head->next = NULL;
> - local_irq_save(oldirq);
> + local_irq_save(flags);
> rdp = RCU_DATA_ME();
> spin_lock(&rdp->lock);
> __rcu_advance_callbacks(rdp);
> @@ -708,7 +1238,7 @@ void fastcall call_rcu_preempt(struct rc
> rdp->nexttail = &head->next;
> RCU_TRACE_RDP(rcupreempt_trace_next_add, rdp);
> spin_unlock(&rdp->lock);
> - local_irq_restore(oldirq);
> + local_irq_restore(flags);
> }
> EXPORT_SYMBOL_GPL(call_rcu_preempt);
>
> @@ -757,6 +1287,11 @@ int rcu_pending_rt(int cpu)
> return 0;
> }
>
> +/*
> + * Initialize RCU. This is called very early in boot, so is restricted
> + * to very simple operations. Don't even think about messing with anything
> + * that involves the scheduler, as it doesn't exist yet.
> + */
> void __init rcu_init_rt(void)
> {
> int cpu;
> @@ -778,6 +1313,7 @@ void __init rcu_init_rt(void)
> rdp->donelist = NULL;
> rdp->donetail = &rdp->donelist;
> }
> + init_rcu_boost_early();
> }
>
> /*
> diff -urpNa -X dontdiff linux-2.6.22-E-hotplug/kernel/rtmutex.c linux-2.6.22-F-boostrcu/kernel/rtmutex.c
> --- linux-2.6.22-E-hotplug/kernel/rtmutex.c 2007-07-08 16:32:17.000000000 -0700
> +++ linux-2.6.22-F-boostrcu/kernel/rtmutex.c 2007-08-24 11:24:59.000000000 -0700
> @@ -111,11 +111,12 @@ static inline void mark_rt_mutex_waiters
> */
> int rt_mutex_getprio(struct task_struct *task)
> {
> + int prio = min(task->normal_prio, get_rcu_prio(task));
> +
> if (likely(!task_has_pi_waiters(task)))
> - return task->normal_prio;
> + return prio;
>
> - return min(task_top_pi_waiter(task)->pi_list_entry.prio,
> - task->normal_prio);
> + return min(task_top_pi_waiter(task)->pi_list_entry.prio, prio);
> }
>
> /*
> diff -urpNa -X dontdiff linux-2.6.22-E-hotplug/kernel/sched.c linux-2.6.22-F-boostrcu/kernel/sched.c
> --- linux-2.6.22-E-hotplug/kernel/sched.c 2007-07-08 16:32:17.000000000 -0700
> +++ linux-2.6.22-F-boostrcu/kernel/sched.c 2007-08-24 11:24:59.000000000 -0700
> @@ -1702,6 +1702,7 @@ void fastcall sched_fork(struct task_str
> * Make sure we do not leak PI boosting priority to the child:
> */
> p->prio = current->normal_prio;
> + set_rcu_prio(p, MAX_PRIO);
>
> INIT_LIST_HEAD(&p->run_list);
> p->array = NULL;
> @@ -1784,6 +1785,7 @@ void fastcall wake_up_new_task(struct ta
> else {
> p->prio = current->prio;
> p->normal_prio = current->normal_prio;
> + set_rcu_prio(p, MAX_PRIO);
> list_add_tail(&p->run_list, ¤t->run_list);
> p->array = current->array;
> p->array->nr_active++;
> @@ -3590,6 +3592,8 @@ asmlinkage void __sched schedule(void)
> }
> profile_hit(SCHED_PROFILING, __builtin_return_address(0));
>
> + rcu_preempt_boost();
> +
> need_resched:
> preempt_disable();
> prev = current;
> @@ -5060,6 +5064,7 @@ void __cpuinit init_idle(struct task_str
> idle->sleep_avg = 0;
> idle->array = NULL;
> idle->prio = idle->normal_prio = MAX_PRIO;
> + set_rcu_prio(idle, MAX_PRIO);
> idle->state = TASK_RUNNING;
> idle->cpus_allowed = cpumask_of_cpu(cpu);
> set_task_cpu(idle, cpu);
> diff -urpNa -X dontdiff linux-2.6.22-E-hotplug/lib/Kconfig.debug linux-2.6.22-F-boostrcu/lib/Kconfig.debug
> --- linux-2.6.22-E-hotplug/lib/Kconfig.debug 2007-07-08 16:32:17.000000000 -0700
> +++ linux-2.6.22-F-boostrcu/lib/Kconfig.debug 2007-08-24 11:24:59.000000000 -0700
> @@ -391,6 +391,40 @@ config RCU_TORTURE_TEST
> Say M if you want the RCU torture tests to build as a module.
> Say N if you are unsure.
>
> +config RCU_TRACE
> + bool "Enable tracing for RCU - currently stats in debugfs"
> + select DEBUG_FS
> + depends on DEBUG_KERNEL
> + default y
> + help
> + This option provides tracing in RCU which presents stats
> + in debugfs for debugging RCU implementation.
> +
> + Say Y here if you want to enable RCU tracing
> + Say N if you are unsure.
> +
> +config PREEMPT_RCU_BOOST_STATS
> + bool "Enable RCU priority-boosting statistic printing"
> + depends on PREEMPT_RCU_BOOST
> + depends on DEBUG_KERNEL
> + default n
> + help
> + This option enables debug printk()s of RCU boost statistics,
> + which are normally only used to debug RCU priority boost
> + implementations.
> +
> + Say N if you are unsure.
> +
> +config PREEMPT_RCU_BOOST_STATS_INTERVAL
> + int "RCU priority-boosting statistic printing interval (seconds)"
> + depends on PREEMPT_RCU_BOOST_STATS
> + default 100
> + range 10 86400
> + help
> + This option controls the timing of debug printk()s of RCU boost
> + statistics, which are normally only used to debug RCU priority
> + boost implementations.
> +
> config LKDTM
> tristate "Linux Kernel Dump Test Tool Module"
> depends on DEBUG_KERNEL
--
Gautham R Shenoy
Linux Technology Center
IBM India.
"Freedom comes with a price tag of responsibility, which is still a bargain,
because Freedom is priceless!"
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[Index of Archives]
[Kernel Newbies]
[Netfilter]
[Bugtraq]
[Photo]
[Stuff]
[Gimp]
[Yosemite News]
[MIPS Linux]
[ARM Linux]
[Linux Security]
[Linux RAID]
[Video 4 Linux]
[Linux for the blind]
[Linux Resources]
