From ca4edb3fc7510c81c54e8dfe26784d3361020a8c Mon Sep 17 00:00:00 2001 From: Glenn Elliott Date: Tue, 26 Nov 2013 14:41:19 -0500 Subject: Add C-FL scheduler plugin. This patch adds a C-FL scheduler plugin. Original work by Jeremy Erikson, port to latest Litmus by Namhoon Kim, and cleanup and commit by Glenn Elliott. --- include/litmus/edf_split_common.h | 25 + include/litmus/litmus.h | 4 + include/litmus/rt_param.h | 6 + litmus/Kconfig | 16 + litmus/Makefile | 1 + litmus/edf_split_common.c | 106 ++++ litmus/sched_cfl_split.c | 1006 +++++++++++++++++++++++++++++++++++++ 7 files changed, 1164 insertions(+) create mode 100644 include/litmus/edf_split_common.h create mode 100644 litmus/edf_split_common.c create mode 100644 litmus/sched_cfl_split.c diff --git a/include/litmus/edf_split_common.h b/include/litmus/edf_split_common.h new file mode 100644 index 000000000000..4e7c0ce23c9d --- /dev/null +++ b/include/litmus/edf_split_common.h @@ -0,0 +1,25 @@ +/* + * EDF common data structures and utility functions shared by all EDF + * based scheduler plugins + */ + +/* CLEANUP: Add comments and make it less messy. + * + */ + +#ifndef __UNC_EDF_SPLIT_COMMON_H__ +#define __UNC_EDF_SPLIT_COMMON_H__ + +#include + +void edf_split_domain_init(rt_domain_t* rt, check_resched_needed_t resched, + release_jobs_t release); + +int edf_split_higher_prio(struct task_struct* first, + struct task_struct* second); + +int edf_split_ready_order(struct bheap_node* a, struct bheap_node* b); + +int edf_split_preemption_needed(rt_domain_t* rt, struct task_struct *t); + +#endif diff --git a/include/litmus/litmus.h b/include/litmus/litmus.h index e35c38c4c0a2..c240d9c07169 100644 --- a/include/litmus/litmus.h +++ b/include/litmus/litmus.h @@ -67,6 +67,7 @@ void litmus_exit_task(struct task_struct *tsk); /* job_param macros */ #define get_exec_time(t) (tsk_rt(t)->job_params.exec_time) #define get_deadline(t) (tsk_rt(t)->job_params.deadline) +#define get_subjob_deadline(t) (tsk_rt(t)->job_params.subjob_deadline) #define get_release(t) (tsk_rt(t)->job_params.release) #define get_lateness(t) (tsk_rt(t)->job_params.lateness) @@ -118,6 +119,9 @@ static inline lt_t litmus_clock(void) #define earlier_release(a, b) (lt_before(\ (a)->rt_param.job_params.release,\ (b)->rt_param.job_params.release)) +#define earlier_subjob_deadline(a, b) (lt_before(\ + (a)->rt_param.job_params.subjob_deadline,\ + (b)->rt_param.job_params.subjob_deadline)) void preempt_if_preemptable(struct task_struct* t, int on_cpu); diff --git a/include/litmus/rt_param.h b/include/litmus/rt_param.h index fe4b31320ac8..6160a1635227 100644 --- a/include/litmus/rt_param.h +++ b/include/litmus/rt_param.h @@ -76,6 +76,7 @@ struct rt_task { lt_t period; lt_t relative_deadline; lt_t phase; + int split; unsigned int cpu; unsigned int priority; task_class_t cls; @@ -149,6 +150,11 @@ struct rt_job { /* What is the current deadline? */ lt_t deadline; +#ifdef CONFIG_JOB_SPLITTING + /* What is the deadline of the current subjob under splitting? */ + lt_t subjob_deadline; +#endif + /* How much service has this job received so far? */ lt_t exec_time; diff --git a/litmus/Kconfig b/litmus/Kconfig index 11f2801a943f..49017b26d0d3 100644 --- a/litmus/Kconfig +++ b/litmus/Kconfig @@ -12,6 +12,15 @@ config PLUGIN_CEDF On smaller platforms (e.g., ARM PB11MPCore), using C-EDF makes little sense since there aren't any shared caches. +config PLUGIN_CFL + bool "Clustered-Fair-Lateness" + depends on X86 && SYSFS && JOB_SPLITTING + default n + help + Include the Clustered Fair Lateness (C-FL) plugin in the kernel. + This implements Anderson and Erickson's EDF-based scheduler. + Supports job splitting. + config PLUGIN_PFAIR bool "PFAIR" depends on HIGH_RES_TIMERS && HZ_PERIODIC && HZ = "1000" @@ -26,6 +35,13 @@ config PLUGIN_PFAIR If unsure, say Yes. +config JOB_SPLITTING + bool "Job Splitting" + default n + help + Enable job-splitting features for fair-lateness schedulers, such + as C-FL. + config RELEASE_MASTER bool "Release-master Support" depends on ARCH_HAS_SEND_PULL_TIMERS && SMP diff --git a/litmus/Makefile b/litmus/Makefile index 52f407cad77c..6c83cf1734ba 100644 --- a/litmus/Makefile +++ b/litmus/Makefile @@ -25,6 +25,7 @@ obj-y = sched_plugin.o litmus.o \ obj-$(CONFIG_PLUGIN_CEDF) += sched_cedf.o obj-$(CONFIG_PLUGIN_PFAIR) += sched_pfair.o +obj-$(CONFIG_PLUGIN_CFL) += sched_cfl_split.o edf_split_common.o obj-$(CONFIG_SCHED_CPU_AFFINITY) += affinity.o obj-$(CONFIG_SCHED_PGM) += pgm.o diff --git a/litmus/edf_split_common.c b/litmus/edf_split_common.c new file mode 100644 index 000000000000..7d7f5429e4df --- /dev/null +++ b/litmus/edf_split_common.c @@ -0,0 +1,106 @@ +/* + * kernel/edf_split_common.c + * + * Common functions for EDF based scheduler with split jobs. + */ + +#include +#include +#include + +#include +#include +#include + +#include +#include + +/* edf_split_higher_prio - returns true if first has a higher subjob + * EDF priority than second. + * + * both first and second may be NULL + */ +int edf_split_higher_prio(struct task_struct* first, + struct task_struct* second) +{ + struct task_struct *first_task = first; + struct task_struct *second_task = second; + + /* There is no point in comparing a task to itself. */ + if (first && first == second) { + TRACE_TASK(first, + "WARNING: pointless edf priority comparison.\n"); + return 0; + } + + /* check for NULL tasks */ + if (!first || !second) + return first && !second; + +#ifdef CONFIG_LITMUS_LOCKING + + /* Check for inherited priorities. Change task + * used for comparison in such a case. + */ + if (unlikely(first->rt_param.inh_task)) + first_task = first->rt_param.inh_task; + if (unlikely(second->rt_param.inh_task)) + second_task = second->rt_param.inh_task; + + /* Check for priority boosting. Tie-break by start of boosting. + */ + if (unlikely(is_priority_boosted(first_task))) { + /* first_task is boosted, how about second_task? */ + if (!is_priority_boosted(second_task) || + lt_before(get_boost_start(first_task), + get_boost_start(second_task))) + return 1; + else + return 0; + } else if (unlikely(is_priority_boosted(second_task))) + /* second_task is boosted, first is not*/ + return 0; +#endif + + if (earlier_subjob_deadline(first_task, second_task)) { + return 1; + } + else if (get_subjob_deadline(first_task) == get_subjob_deadline(second_task)) { + /* use normal edf to tie-break */ + return edf_higher_prio(first, second); + } + return 0; /* fall-through. prio(second_task) > prio(first_task) */ +} + +int edf_split_ready_order(struct bheap_node* a, struct bheap_node* b) +{ + return edf_split_higher_prio(bheap2task(a), bheap2task(b)); +} + +void edf_split_domain_init(rt_domain_t* rt, check_resched_needed_t resched, + release_jobs_t release) +{ + rt_domain_init(rt, edf_split_ready_order, resched, release); +} + +/* need_to_preempt - check whether the task t needs to be preempted + * call only with irqs disabled and with ready_lock acquired + * THIS DOES NOT TAKE NON-PREEMPTIVE SECTIONS INTO ACCOUNT! + */ +int edf_split_preemption_needed(rt_domain_t* rt, struct task_struct *t) +{ + /* we need the read lock for edf_ready_queue */ + /* no need to preempt if there is nothing pending */ + if (!__jobs_pending(rt)) + return 0; + /* we need to reschedule if t doesn't exist */ + if (!t) + return 1; + + /* NOTE: We cannot check for non-preemptibility since we + * don't know what address space we're currently in. + */ + + /* make sure to get non-rt stuff out of the way */ + return !is_realtime(t) || edf_split_higher_prio(__next_ready(rt), t); +} diff --git a/litmus/sched_cfl_split.c b/litmus/sched_cfl_split.c new file mode 100644 index 000000000000..7d9302eb296b --- /dev/null +++ b/litmus/sched_cfl_split.c @@ -0,0 +1,1006 @@ +/* + * litmus/sched_cfl_split.c + * + * Implementation of a clustered version of the C-FL scheduling algorithm, + * with job splitting. + * + * This implementation is based on C-FL-split: + * - CPUs are clustered around L2 or L3 caches. + * - Clusters topology is automatically detected (this is arch dependent + * and is working only on x86 at the moment --- and only with modern + * cpus that exports cpuid4 information) + * - The plugins _does not_ attempt to put tasks in the right cluster i.e. + * the programmer needs to be aware of the topology to place tasks + * in the desired cluster + * - default clustering is around L2 cache (cache index = 2) + * supported clusters are: L1 (private cache: pedf), L2, L3, ALL (all + * online_cpus are placed in a single cluster). + * + * For details on functions, take a look at sched_gsn_edf.c + * + * Currently, we do not support changes in the number of online cpus. + * If the num_online_cpus() dynamically changes, the plugin is broken. + * + * This version uses the simple approach and serializes all scheduling + * decisions by the use of a queue lock. This is probably not the + * best way to do it, but it should suffice for now. + */ + +#include +#include +#include +#include + +#include + +#include +#include +#include +#include +#include +#include +#include + +#include + +#include + +#ifdef CONFIG_SCHED_CPU_AFFINITY +#include +#endif + +/* to configure the cluster size */ +#include +#include + +/* Reference configuration variable. Determines which cache level is used to + * group CPUs into clusters. GLOBAL_CLUSTER, which is the default, means that + * all CPUs form a single cluster (just like G-FL). + */ +static enum cache_level cluster_config = GLOBAL_CLUSTER; + +struct clusterdomain; + +/* cpu_entry_t - maintain the linked and scheduled state + * + * A cpu also contains a pointer to the cflsplit_domain_t cluster + * that owns it (struct clusterdomain*) + */ +typedef struct { + int cpu; + struct clusterdomain* cluster; /* owning cluster */ + struct task_struct* linked; /* only RT tasks */ + struct task_struct* scheduled; /* only RT tasks */ + atomic_t will_schedule; /* prevent unneeded IPIs */ + struct bheap_node* hn; + struct hrtimer split_timer; + int timer_armed; +} cpu_entry_t; + +/* one cpu_entry_t per CPU */ +DEFINE_PER_CPU(cpu_entry_t, cflsplit_cpu_entries); + +#define set_will_schedule() \ + (atomic_set(&__get_cpu_var(cflsplit_cpu_entries).will_schedule, 1)) +#define clear_will_schedule() \ + (atomic_set(&__get_cpu_var(cflsplit_cpu_entries).will_schedule, 0)) +#define test_will_schedule(cpu) \ + (atomic_read(&per_cpu(cflsplit_cpu_entries, cpu).will_schedule)) + +/* + * In C-FL-split there is a cflsplit domain _per_ cluster + * The number of clusters is dynamically determined accordingly to the + * total cpu number and the cluster size + */ +typedef struct clusterdomain { + /* rt_domain for this cluster */ + rt_domain_t domain; + /* cpus in this cluster */ + cpu_entry_t* *cpus; + /* map of this cluster cpus */ + cpumask_var_t cpu_map; + /* the cpus queue themselves according to priority in here */ + struct bheap_node *heap_node; + struct bheap cpu_heap; + /* lock for this cluster */ +#define cluster_lock domain.ready_lock +} cflsplit_domain_t; + +/* a cflsplit_domain per cluster; allocation is done at init/activation time */ +cflsplit_domain_t *cflsplit; + +#define remote_cluster(cpu) ((cflsplit_domain_t *) per_cpu(cflsplit_cpu_entries, cpu).cluster) +#define task_cpu_cluster(task) remote_cluster(get_partition(task)) + +/* Uncomment WANT_ALL_SCHED_EVENTS if you want to see all scheduling + * decisions in the TRACE() log; uncomment VERBOSE_INIT for verbose + * information during the initialization of the plugin (e.g., topology) +#define WANT_ALL_SCHED_EVENTS + */ +#define VERBOSE_INIT + +inline static int get_slice_num(struct task_struct* t) +{ + int basic = ((t->rt_param.job_params.exec_time * + t->rt_param.task_params.split) / + t->rt_param.task_params.exec_cost) + 1; + if (basic <= t->rt_param.task_params.split){ + return basic; + } + else{ + /*Since we don't police budget, just leave where it's at.*/ + return t->rt_param.task_params.split; + } +} + +/* Returns the appropriate subjob deadline.*/ +inline static lt_t get_proper_deadline(struct task_struct* t) +{ + unsigned int num_cpus = num_online_cpus(); + return t->rt_param.job_params.release + + ((t->rt_param.task_params.period * get_slice_num(t)) + / t->rt_param.task_params.split) + /* G-FL correction */ + - (((num_cpus - 1) * t->rt_param.task_params.exec_cost) + / (num_cpus * t->rt_param.task_params.split)); +} + +/* Tells us if the current deadline is too small.*/ +inline static int needs_deadline_move(struct task_struct* t) +{ + BUG_ON(get_proper_deadline(t) < t->rt_param.job_params.subjob_deadline); + return get_proper_deadline(t) != tsk_rt(t)->job_params.subjob_deadline; +} + +/*Returns execution time until the next deadline move. + * 0 means the task has no more deadline moves + */ +inline static lt_t time_to_next_move(struct task_struct* t) +{ + if (get_slice_num(t) == t->rt_param.task_params.split){ + return 0; + } + /* +1 upper bounds ceiling, since integer division is floor*/ + return ((get_slice_num(t) * t->rt_param.task_params.exec_cost) + / t->rt_param.task_params.split) + 1 + - t->rt_param.job_params.exec_time; +} + +/* Timer stuff - similar to budget.c. */ +static enum hrtimer_restart on_split_timeout(struct hrtimer *timer) +{ + cpu_entry_t* st = container_of(timer, + cpu_entry_t, + split_timer); + + unsigned long flags; + + local_irq_save(flags); + TRACE("split timer fired: %llu\n", litmus_clock()); + st->timer_armed = 0; + /* Activate scheduler */ + litmus_reschedule_local(); + local_irq_restore(flags); + + return HRTIMER_NORESTART; +} + +static void cancel_split_timer(cpu_entry_t* ce) +{ + int ret; + + TRACE("cancelling split time.\n"); + + /* Since interrupts are disabled and et->timer_armed is only + * modified locally, we do not need any locks. + */ + + if (ce->timer_armed) { + ret = hrtimer_try_to_cancel(&ce->split_timer); + /* Should never be inactive. */ + BUG_ON(ret == 0); + /* Should never be running concurrently.*/ + BUG_ON(ret == -1); + + ce->timer_armed = 0; + } +} + +/* assumes called with IRQs off */ +static void arm_split_timer(cpu_entry_t *ce, + struct task_struct* t) +{ + lt_t when_to_fire; + lt_t time_to_move; + lt_t now = litmus_clock(); + + /* __hrtimer_start_range_ns() cancels the timer + * anyway, so we don't have to check whether it is still armed */ + + /*We won't do any new deadline moves if the budget has been exhausted*/ + if (likely(!is_np(t) && (time_to_move = time_to_next_move(t)))) { + when_to_fire = now + time_to_move; + TRACE_TASK(t, "actually arming for %llu into the future\n", + time_to_move); + __hrtimer_start_range_ns(&ce->split_timer, + ns_to_ktime(when_to_fire), + 0 /* delta */, + HRTIMER_MODE_ABS_PINNED, + 0 /* no wakeup */); + ce->timer_armed = 1; + } +} + +static int cpu_lower_prio(struct bheap_node *_a, struct bheap_node *_b) +{ + cpu_entry_t *a, *b; + a = _a->value; + b = _b->value; + /* Note that a and b are inverted: we want the lowest-priority CPU at + * the top of the heap. + */ + return edf_split_higher_prio(b->linked, a->linked); +} + +/* update_cpu_position - Move the cpu entry to the correct place to maintain + * order in the cpu queue. Caller must hold cflsplit lock. + */ +static void update_cpu_position(cpu_entry_t *entry) +{ + cflsplit_domain_t *cluster = entry->cluster; + + if (likely(bheap_node_in_heap(entry->hn))) + bheap_delete(cpu_lower_prio, + &cluster->cpu_heap, + entry->hn); + + bheap_insert(cpu_lower_prio, &cluster->cpu_heap, entry->hn); +} + +/* caller must hold cflsplit lock */ +static cpu_entry_t* lowest_prio_cpu(cflsplit_domain_t *cluster) +{ + struct bheap_node* hn; + hn = bheap_peek(cpu_lower_prio, &cluster->cpu_heap); + return hn->value; +} + + +/* link_task_to_cpu - Update the link of a CPU. + * Handles the case where the to-be-linked task is already + * scheduled on a different CPU. + */ +static noinline void link_task_to_cpu(struct task_struct* linked, + cpu_entry_t *entry) +{ + cpu_entry_t *sched; + struct task_struct* tmp; + int on_cpu; + + BUG_ON(linked && !is_realtime(linked)); + + /* Currently linked task is set to be unlinked. */ + if (entry->linked) { + entry->linked->rt_param.linked_on = NO_CPU; + } + + /* Link new task to CPU. */ + if (linked) { + /* handle task is already scheduled somewhere! */ + on_cpu = linked->rt_param.scheduled_on; + if (on_cpu != NO_CPU) { + sched = &per_cpu(cflsplit_cpu_entries, on_cpu); + /* this should only happen if not linked already */ + BUG_ON(sched->linked == linked); + + /* If we are already scheduled on the CPU to which we + * wanted to link, we don't need to do the swap -- + * we just link ourselves to the CPU and depend on + * the caller to get things right. + */ + if (entry != sched) { + TRACE_TASK(linked, + "already scheduled on %d, updating link.\n", + sched->cpu); + tmp = sched->linked; + linked->rt_param.linked_on = sched->cpu; + sched->linked = linked; + update_cpu_position(sched); + linked = tmp; + } + } + if (linked) /* might be NULL due to swap */ + linked->rt_param.linked_on = entry->cpu; + } + entry->linked = linked; +#ifdef WANT_ALL_SCHED_EVENTS + if (linked) + TRACE_TASK(linked, "linked to %d.\n", entry->cpu); + else + TRACE("NULL linked to %d.\n", entry->cpu); +#endif + update_cpu_position(entry); +} + +/* unlink - Make sure a task is not linked any longer to an entry + * where it was linked before. Must hold cflsplit_lock. + */ +static noinline void unlink(struct task_struct* t) +{ + cpu_entry_t *entry; + + if (t->rt_param.linked_on != NO_CPU) { + /* unlink */ + entry = &per_cpu(cflsplit_cpu_entries, t->rt_param.linked_on); + t->rt_param.linked_on = NO_CPU; + link_task_to_cpu(NULL, entry); + } else if (is_queued(t)) { + /* This is an interesting situation: t is scheduled, + * but was just recently unlinked. It cannot be + * linked anywhere else (because then it would have + * been relinked to this CPU), thus it must be in some + * queue. We must remove it from the list in this + * case. + * + * in C-FL-split case is should be somewhere in the queue for + * its domain, therefore and we can get the domain using + * task_cpu_cluster + */ + remove(&(task_cpu_cluster(t))->domain, t); + } +} + + +/* preempt - force a CPU to reschedule + */ +static void preempt(cpu_entry_t *entry) +{ + preempt_if_preemptable(entry->scheduled, entry->cpu); +} + +/* requeue - Put an unlinked task into gsn-edf domain. + * Caller must hold cflsplit_lock. + */ +static noinline void requeue(struct task_struct* task) +{ + cflsplit_domain_t *cluster = task_cpu_cluster(task); + BUG_ON(!task); + /* sanity check before insertion */ + BUG_ON(is_queued(task)); + + if (is_early_releasing(task) || is_released(task, litmus_clock())) + __add_ready(&cluster->domain, task); + else { + /* it has got to wait */ + add_release(&cluster->domain, task); + } +} + +#ifdef CONFIG_SCHED_CPU_AFFINITY +static cpu_entry_t* cflsplit_get_nearest_available_cpu( + cflsplit_domain_t *cluster, cpu_entry_t *start) +{ + cpu_entry_t *affinity; + + get_nearest_available_cpu(affinity, start, cflsplit_cpu_entries, +#ifdef CONFIG_RELEASE_MASTER + cluster->domain.release_master +#else + NO_CPU +#endif + ); + + /* make sure CPU is in our cluster */ + if (affinity && cpu_isset(affinity->cpu, *cluster->cpu_map)) + return(affinity); + else + return(NULL); +} +#endif + + +/* check for any necessary preemptions */ +static void check_for_preemptions(cflsplit_domain_t *cluster) +{ + struct task_struct *task; + cpu_entry_t *last; + + for(last = lowest_prio_cpu(cluster); + edf_split_preemption_needed(&cluster->domain, last->linked); + last = lowest_prio_cpu(cluster)) { + /* preemption necessary */ + task = __take_ready(&cluster->domain); + TRACE("check_for_preemptions: attempting to link task %d to %d\n", + task->pid, last->cpu); +#ifdef CONFIG_SCHED_CPU_AFFINITY + { + cpu_entry_t *affinity = + cflsplit_get_nearest_available_cpu(cluster, + &per_cpu(cflsplit_cpu_entries, task_cpu(task))); + if(affinity) + last = affinity; + else if(requeue_preempted_job(last->linked)) + requeue(last->linked); + } +#else + if (requeue_preempted_job(last->linked)) + requeue(last->linked); +#endif + link_task_to_cpu(task, last); + preempt(last); + } +} + +/* cflsplit_job_arrival: task is either resumed or released */ +static noinline void cflsplit_job_arrival(struct task_struct* task) +{ + cflsplit_domain_t *cluster = task_cpu_cluster(task); + BUG_ON(!task); + + requeue(task); + check_for_preemptions(cluster); +} + +static void cflsplit_release_jobs(rt_domain_t* rt, struct bheap* tasks) +{ + cflsplit_domain_t* cluster = container_of(rt, cflsplit_domain_t, domain); + unsigned long flags; + + raw_spin_lock_irqsave(&cluster->cluster_lock, flags); + + __merge_ready(&cluster->domain, tasks); + check_for_preemptions(cluster); + + raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags); +} + +/* caller holds cflsplit_lock */ +static noinline void job_completion(struct task_struct *t, int forced) +{ + BUG_ON(!t); + + sched_trace_task_completion(t, forced); + + TRACE_TASK(t, "job_completion().\n"); + + /* set flags */ + tsk_rt(t)->completed = 0; + /* prepare for next period */ + prepare_for_next_period(t); + /* We now also set the subjob deadline to what it should be for + * scheduling priority. + */ + t->rt_param.job_params.subjob_deadline = get_proper_deadline(t); + if (is_early_releasing(t) || is_released(t, litmus_clock())) + sched_trace_task_release(t); + /* unlink */ + unlink(t); + /* requeue + * But don't requeue a blocking task. */ + if (is_running(t)) + cflsplit_job_arrival(t); +} + +static void move_deadline(struct task_struct *t) +{ + tsk_rt(t)->job_params.subjob_deadline = get_proper_deadline(t); + /* Check if rescheduling needed with lower priority. */ + unlink(t); + cflsplit_job_arrival(t); +} + +/* cflsplit_tick - this function is called for every local timer + * interrupt. + * + * checks whether the current task has expired and checks + * whether we need to preempt it if it has not expired + */ +static void cflsplit_tick(struct task_struct* t) +{ + if (is_realtime(t) && budget_enforced(t) && budget_exhausted(t)) { + if (!is_np(t)) { + /* np tasks will be preempted when they become + * preemptable again + */ + litmus_reschedule_local(); + set_will_schedule(); + TRACE("cflsplit_scheduler_tick: " + "%d is preemptable " + " => FORCE_RESCHED\n", t->pid); + } else if (is_user_np(t)) { + TRACE("cflsplit_scheduler_tick: " + "%d is non-preemptable, " + "preemption delayed.\n", t->pid); + request_exit_np(t); + } + } +} + +/* Getting schedule() right is a bit tricky. schedule() may not make any + * assumptions on the state of the current task since it may be called for a + * number of reasons. The reasons include a scheduler_tick() determined that it + * was necessary, because sys_exit_np() was called, because some Linux + * subsystem determined so, or even (in the worst case) because there is a bug + * hidden somewhere. Thus, we must take extreme care to determine what the + * current state is. + * + * The CPU could currently be scheduling a task (or not), be linked (or not). + * + * The following assertions for the scheduled task could hold: + * + * - !is_running(scheduled) // the job blocks + * - scheduled->timeslice == 0 // the job completed (forcefully) + * - is_completed() // the job completed (by syscall) + * - linked != scheduled // we need to reschedule (for any reason) + * - is_np(scheduled) // rescheduling must be delayed, + * sys_exit_np must be requested + * + * Any of these can occur together. + */ +static struct task_struct* cflsplit_schedule(struct task_struct * prev) +{ + cpu_entry_t* entry = &__get_cpu_var(cflsplit_cpu_entries); + cflsplit_domain_t *cluster = entry->cluster; + int out_of_time, sleep, preempt, np, exists, blocks, needs_move; + struct task_struct* next = NULL; + +#ifdef CONFIG_RELEASE_MASTER + /* Bail out early if we are the release master. + * The release master never schedules any real-time tasks. + */ + if (unlikely(cluster->domain.release_master == entry->cpu)) { + sched_state_task_picked(); + return NULL; + } +#endif + + raw_spin_lock(&cluster->cluster_lock); + clear_will_schedule(); + + /* sanity checking */ + BUG_ON(entry->scheduled && entry->scheduled != prev); + BUG_ON(entry->scheduled && !is_realtime(prev)); + BUG_ON(is_realtime(prev) && !entry->scheduled); + + /* (0) Determine state */ + exists = entry->scheduled != NULL; + blocks = exists && !is_running(entry->scheduled); + out_of_time = exists && + budget_enforced(entry->scheduled) && + budget_exhausted(entry->scheduled); + needs_move = exists && needs_deadline_move(entry->scheduled); + np = exists && is_np(entry->scheduled); + sleep = exists && is_completed(entry->scheduled); + preempt = entry->scheduled != entry->linked; + +#ifdef WANT_ALL_SCHED_EVENTS + TRACE_TASK(prev, "invoked cflsplit_schedule.\n"); +#endif + + if (exists) + TRACE_TASK(prev, + "blocks:%d out_of_time:%d needs_move: %d np:%d" + " sleep:%d preempt:%d state:%d sig:%d\n", + blocks, out_of_time, needs_move, np, sleep, preempt, + prev->state, signal_pending(prev)); + if (entry->linked && preempt) + TRACE_TASK(prev, "will be preempted by %s/%d\n", + entry->linked->comm, entry->linked->pid); + + + /* If a task blocks we have no choice but to reschedule. + */ + if (blocks) + unlink(entry->scheduled); + + /* Request a sys_exit_np() call if we would like to preempt but cannot. + * We need to make sure to update the link structure anyway in case + * that we are still linked. Multiple calls to request_exit_np() don't + * hurt. + * + * Job deadline moves handled similarly + */ + if (np && (out_of_time || preempt || sleep)) { + unlink(entry->scheduled); + request_exit_np(entry->scheduled); + } + else if (np && needs_move) { + request_exit_np(entry->scheduled); + } + + /* Any task that is preemptable and either exhausts its execution + * budget or wants to sleep completes. We may have to reschedule after + * this. Don't do a job completion if we block (can't have timers running + * for blocked jobs). Preemption go first for the same reason. + */ + if (!np && (out_of_time || sleep) && !blocks) + job_completion(entry->scheduled, !sleep); + else if (!np && needs_move && !blocks) { + move_deadline(entry->scheduled); + } + + /* Link pending task if we became unlinked. + */ + if (!entry->linked) + link_task_to_cpu(__take_ready(&cluster->domain), entry); + + /* The final scheduling decision. Do we need to switch for some reason? + * If linked is different from scheduled, then select linked as next. + */ + if ((!np || blocks) && + entry->linked != entry->scheduled) { + /* Schedule a linked job? */ + if (entry->linked) { + entry->linked->rt_param.scheduled_on = entry->cpu; + next = entry->linked; + } + if (entry->scheduled) { + /* not gonna be scheduled soon */ + entry->scheduled->rt_param.scheduled_on = NO_CPU; + TRACE_TASK(entry->scheduled, "scheduled_on = NO_CPU\n"); + } + } else + /* Only override Linux scheduler if we have a real-time task + * scheduled that needs to continue. + */ + if (exists) + next = prev; + + sched_state_task_picked(); + raw_spin_unlock(&cluster->cluster_lock); + + if (next) { + arm_split_timer(entry, next); + } + else if (entry->timer_armed) { + cancel_split_timer(entry); + } + +#ifdef WANT_ALL_SCHED_EVENTS + TRACE("cflsplit_lock released, next=0x%p\n", next); + + if (next) + TRACE_TASK(next, "scheduled at %llu\n", litmus_clock()); + else if (exists && !next) + TRACE("becomes idle at %llu.\n", litmus_clock()); +#endif + + + return next; +} + + +/* _finish_switch - we just finished the switch away from prev + */ +static void cflsplit_finish_switch(struct task_struct *prev) +{ + cpu_entry_t* entry = &__get_cpu_var(cflsplit_cpu_entries); + + entry->scheduled = is_realtime(current) ? current : NULL; +#ifdef WANT_ALL_SCHED_EVENTS + TRACE_TASK(prev, "switched away from\n"); +#endif +} + + +static void cflsplit_release_at(struct task_struct *t, lt_t start) +{ + release_at(t, start); + t->rt_param.job_params.subjob_deadline = get_proper_deadline(t); +} + + +/* Prepare a task for running in RT mode + */ +static void cflsplit_task_new(struct task_struct * t, int on_rq, int is_scheduled) +{ + unsigned long flags; + cpu_entry_t* entry; + cflsplit_domain_t* cluster; + + TRACE("gsn edf: task new %d\n", t->pid); + + /* the cluster doesn't change even if t is scheduled */ + cluster = task_cpu_cluster(t); + + raw_spin_lock_irqsave(&cluster->cluster_lock, flags); + + /* setup job params */ + cflsplit_release_at(t, litmus_clock()); + + if (is_scheduled) { + entry = &per_cpu(cflsplit_cpu_entries, task_cpu(t)); + BUG_ON(entry->scheduled); + +#ifdef CONFIG_RELEASE_MASTER + if (entry->cpu != cluster->domain.release_master) { +#endif + entry->scheduled = t; + tsk_rt(t)->scheduled_on = task_cpu(t); +#ifdef CONFIG_RELEASE_MASTER + } else { + /* do not schedule on release master */ + preempt(entry); /* force resched */ + tsk_rt(t)->scheduled_on = NO_CPU; + } +#endif + } else { + t->rt_param.scheduled_on = NO_CPU; + } + t->rt_param.linked_on = NO_CPU; + + if (is_running(t)) + cflsplit_job_arrival(t); + raw_spin_unlock_irqrestore(&(cluster->cluster_lock), flags); +} + +static void cflsplit_task_wake_up(struct task_struct *task) +{ + unsigned long flags; + lt_t now; + cflsplit_domain_t *cluster; + + TRACE_TASK(task, "wake_up at %llu\n", litmus_clock()); + + cluster = task_cpu_cluster(task); + + raw_spin_lock_irqsave(&cluster->cluster_lock, flags); + now = litmus_clock(); + if (is_sporadic(task) && is_tardy(task, now)) { + /* new sporadic release */ + cflsplit_release_at(task, now); + sched_trace_task_release(task); + } + cflsplit_job_arrival(task); + raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags); +} + +static void cflsplit_task_block(struct task_struct *t) +{ + unsigned long flags; + cflsplit_domain_t *cluster; + + TRACE_TASK(t, "block at %llu\n", litmus_clock()); + + cluster = task_cpu_cluster(t); + + /* unlink if necessary */ + raw_spin_lock_irqsave(&cluster->cluster_lock, flags); + unlink(t); + raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags); + + BUG_ON(!is_realtime(t)); +} + + +static void cflsplit_task_exit(struct task_struct * t) +{ + unsigned long flags; + cflsplit_domain_t *cluster = task_cpu_cluster(t); + + /* unlink if necessary */ + raw_spin_lock_irqsave(&cluster->cluster_lock, flags); + unlink(t); + if (tsk_rt(t)->scheduled_on != NO_CPU) { + cpu_entry_t *cpu; + cpu = &per_cpu(cflsplit_cpu_entries, tsk_rt(t)->scheduled_on); + cpu->scheduled = NULL; + tsk_rt(t)->scheduled_on = NO_CPU; + } + raw_spin_unlock_irqrestore(&cluster->cluster_lock, flags); + + BUG_ON(!is_realtime(t)); + TRACE_TASK(t, "RIP\n"); +} + +static long cflsplit_admit_task(struct task_struct* tsk) +{ + return (remote_cluster(task_cpu(tsk)) == task_cpu_cluster(tsk)) ? + 0 : -EINVAL; +} + +/* total number of cluster */ +static int num_clusters; +/* we do not support cluster of different sizes */ +static unsigned int cluster_size; + +#ifdef VERBOSE_INIT +static void print_cluster_topology(cpumask_var_t mask, int cpu) +{ + int chk; + char buf[255]; + + chk = cpulist_scnprintf(buf, 254, mask); + buf[chk] = '\0'; + printk(KERN_INFO "CPU = %d, shared cpu(s) = %s\n", cpu, buf); + +} +#endif + +static int clusters_allocated = 0; + +static void cleanup_cflsplit(void) +{ + int i; + + if (clusters_allocated) { + for (i = 0; i < num_clusters; i++) { + kfree(cflsplit[i].cpus); + kfree(cflsplit[i].heap_node); + free_cpumask_var(cflsplit[i].cpu_map); + } + + kfree(cflsplit); + } +} + +static long cflsplit_activate_plugin(void) +{ + int i, j, cpu, ccpu, cpu_count; + cpu_entry_t *entry; + + cpumask_var_t mask; + int chk = 0; + + /* de-allocate old clusters, if any */ + cleanup_cflsplit(); + + printk(KERN_INFO "C-FL-split: Activate Plugin, cluster configuration = %d\n", + cluster_config); + + /* need to get cluster_size first */ + if(!zalloc_cpumask_var(&mask, GFP_ATOMIC)) + return -ENOMEM; + + if (unlikely(cluster_config == GLOBAL_CLUSTER)) { + cluster_size = num_online_cpus(); + } else { + chk = get_shared_cpu_map(mask, 0, cluster_config); + if (chk) { + /* if chk != 0 then it is the max allowed index */ + printk(KERN_INFO "C-FL-split: Cluster configuration = %d " + "is not supported on this hardware.\n", + cluster_config); + /* User should notice that the configuration failed, so + * let's bail out. */ + return -EINVAL; + } + + cluster_size = cpumask_weight(mask); + } + + if ((num_online_cpus() % cluster_size) != 0) { + /* this can't be right, some cpus are left out */ + printk(KERN_ERR "C-FL-split: Trying to group %d cpus in %d!\n", + num_online_cpus(), cluster_size); + return -1; + } + + num_clusters = num_online_cpus() / cluster_size; + printk(KERN_INFO "C-FL-split: %d cluster(s) of size = %d\n", + num_clusters, cluster_size); + + /* initialize clusters */ + cflsplit = kmalloc(num_clusters * sizeof(cflsplit_domain_t), GFP_ATOMIC); + for (i = 0; i < num_clusters; i++) { + + cflsplit[i].cpus = kmalloc(cluster_size * sizeof(cpu_entry_t), + GFP_ATOMIC); + cflsplit[i].heap_node = kmalloc( + cluster_size * sizeof(struct bheap_node), + GFP_ATOMIC); + bheap_init(&(cflsplit[i].cpu_heap)); + edf_split_domain_init(&(cflsplit[i].domain), NULL, + cflsplit_release_jobs); + + if(!zalloc_cpumask_var(&cflsplit[i].cpu_map, GFP_ATOMIC)) + return -ENOMEM; +#ifdef CONFIG_RELEASE_MASTER + cflsplit[i].domain.release_master = atomic_read(&release_master_cpu); +#endif + } + + /* cycle through cluster and add cpus to them */ + for (i = 0; i < num_clusters; i++) { + + for_each_online_cpu(cpu) { + /* check if the cpu is already in a cluster */ + for (j = 0; j < num_clusters; j++) + if (cpumask_test_cpu(cpu, cflsplit[j].cpu_map)) + break; + /* if it is in a cluster go to next cpu */ + if (j < num_clusters && + cpumask_test_cpu(cpu, cflsplit[j].cpu_map)) + continue; + + /* this cpu isn't in any cluster */ + /* get the shared cpus */ + if (unlikely(cluster_config == GLOBAL_CLUSTER)) + cpumask_copy(mask, cpu_online_mask); + else + get_shared_cpu_map(mask, cpu, cluster_config); + + cpumask_copy(cflsplit[i].cpu_map, mask); +#ifdef VERBOSE_INIT + print_cluster_topology(mask, cpu); +#endif + /* add cpus to current cluster and init cpu_entry_t */ + cpu_count = 0; + for_each_cpu(ccpu, cflsplit[i].cpu_map) { + + entry = &per_cpu(cflsplit_cpu_entries, ccpu); + cflsplit[i].cpus[cpu_count] = entry; + atomic_set(&entry->will_schedule, 0); + entry->cpu = ccpu; + entry->cluster = &cflsplit[i]; + entry->hn = &(cflsplit[i].heap_node[cpu_count]); + hrtimer_init(&entry->split_timer, + CLOCK_MONOTONIC, + HRTIMER_MODE_ABS); + entry->split_timer.function = on_split_timeout; + bheap_node_init(&entry->hn, entry); + + cpu_count++; + + entry->linked = NULL; + entry->scheduled = NULL; +#ifdef CONFIG_RELEASE_MASTER + /* only add CPUs that should schedule jobs */ + if (entry->cpu != entry->cluster->domain.release_master) +#endif + update_cpu_position(entry); + } + /* done with this cluster */ + break; + } + } + + free_cpumask_var(mask); + clusters_allocated = 1; + return 0; +} + +/* Plugin object */ +static struct sched_plugin cflsplit_plugin __cacheline_aligned_in_smp = { + .plugin_name = "C-FL-split", + .finish_switch = cflsplit_finish_switch, + .tick = cflsplit_tick, + .task_new = cflsplit_task_new, + .complete_job = complete_job, + .task_exit = cflsplit_task_exit, + .schedule = cflsplit_schedule, + .release_at = cflsplit_release_at, + .task_wake_up = cflsplit_task_wake_up, + .task_block = cflsplit_task_block, + .admit_task = cflsplit_admit_task, + .activate_plugin = cflsplit_activate_plugin, +}; + +static struct proc_dir_entry *cluster_file = NULL, *cflsplit_dir = NULL; + +static int __init init_cflsplit(void) +{ + int err, fs; + + err = register_sched_plugin(&cflsplit_plugin); + if (!err) { + fs = make_plugin_proc_dir(&cflsplit_plugin, &cflsplit_dir); + if (!fs) + cluster_file = create_cluster_file(cflsplit_dir, &cluster_config); + else + printk(KERN_ERR "Could not allocate C-FL-split procfs dir.\n"); + } + return err; +} + +static void clean_cflsplit(void) +{ + cleanup_cflsplit(); + if (cluster_file) + remove_proc_entry("cluster", cflsplit_dir); + if (cflsplit_dir) + remove_plugin_proc_dir(&cflsplit_plugin); +} + +module_init(init_cflsplit); +module_exit(clean_cflsplit); -- cgit v1.2.2