[PATCH 2/2] sched: Use a 2-d bitmap for searching lowest-pri CPU

From: Gregory Haskins
Date: Mon Apr 21 2008 - 14:43:56 EST


The current code use a linear algorithm which causes scaling issues
on larger SMP machines. This patch replaces that algorithm with a
2-dimensional bitmap to reduce latencies in the wake-up path.

Signed-off-by: Gregory Haskins <ghaskins@xxxxxxxxxx>
---

kernel/Makefile | 1
kernel/sched.c | 7 ++
kernel/sched_cpupri.c | 174 +++++++++++++++++++++++++++++++++++++++++++++++++
kernel/sched_cpupri.h | 36 ++++++++++
kernel/sched_rt.c | 98 ++++++----------------------
5 files changed, 239 insertions(+), 77 deletions(-)

diff --git a/kernel/Makefile b/kernel/Makefile
index 7071d44..8c55d9e 100644
--- a/kernel/Makefile
+++ b/kernel/Makefile
@@ -79,6 +79,7 @@ obj-$(CONFIG_MARKERS) += marker.o
obj-$(CONFIG_LATENCYTOP) += latencytop.o
obj-$(CONFIG_FTRACE) += trace/
obj-$(CONFIG_TRACING) += trace/
+obj-$(CONFIG_SMP) += sched_cpupri.o

ifneq ($(CONFIG_SCHED_NO_NO_OMIT_FRAME_POINTER),y)
# According to Alan Modra <alan@xxxxxxxxxxxxxxxx>, the -fno-omit-frame-pointer is
diff --git a/kernel/sched.c b/kernel/sched.c
index d859649..5899c29 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -75,6 +75,8 @@
#include <asm/tlb.h>
#include <asm/irq_regs.h>

+#include "sched_cpupri.h"
+
/*
* Scheduler clock - returns current time in nanosec units.
* This is default implementation.
@@ -501,6 +503,9 @@ struct root_domain {
*/
cpumask_t rto_mask;
atomic_t rto_count;
+#ifdef CONFIG_SMP
+ struct cpupri cpupri;
+#endif
};

/*
@@ -6923,6 +6928,8 @@ static void init_rootdomain(struct root_domain *rd)

cpus_clear(rd->span);
cpus_clear(rd->online);
+
+ cpupri_init(&rd->cpupri);
}

static void init_defrootdomain(void)
diff --git a/kernel/sched_cpupri.c b/kernel/sched_cpupri.c
new file mode 100644
index 0000000..52154fe
--- /dev/null
+++ b/kernel/sched_cpupri.c
@@ -0,0 +1,174 @@
+/*
+ * kernel/sched_cpupri.c
+ *
+ * CPU priority management
+ *
+ * Copyright (C) 2007-2008 Novell
+ *
+ * Author: Gregory Haskins <ghaskins@xxxxxxxxxx>
+ *
+ * This code tracks the priority of each CPU so that global migration
+ * decisions are easy to calculate. Each CPU can be in a state as follows:
+ *
+ * (INVALID), IDLE, NORMAL, RT1, ... RT99
+ *
+ * going from the lowest priority to the highest. CPUs in the INVALID state
+ * are not eligible for routing. The system maintains this state with
+ * a 2 dimensional bitmap (the first for priority class, the second for cpus
+ * in that class). Therefore a typical application without affinity
+ * restrictions can find a suitable CPU with O(1) complexity (e.g. two bit
+ * searches). For tasks with affinity restrictions, the algorithm has a
+ * worst case complexity of O(min(102, nr_domcpus)), though the scenario that
+ * yields the worst case search is fairly contrived.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; version 2
+ * of the License.
+ */
+
+#include "sched_cpupri.h"
+
+/* Convert between a 140 based task->prio, and our 102 based cpupri */
+static int convert_prio(int prio)
+{
+ int cpupri;
+
+ if (prio == CPUPRI_INVALID)
+ cpupri = CPUPRI_INVALID;
+ else if (prio == MAX_PRIO)
+ cpupri = CPUPRI_IDLE;
+ else if (prio >= MAX_RT_PRIO)
+ cpupri = CPUPRI_NORMAL;
+ else
+ cpupri = MAX_RT_PRIO - prio + 1;
+
+ return cpupri;
+}
+
+#define for_each_cpupri_active(array, idx) \
+ for (idx = find_first_bit(array, CPUPRI_NR_PRIORITIES); \
+ idx < CPUPRI_NR_PRIORITIES; \
+ idx = find_next_bit(array, CPUPRI_NR_PRIORITIES, idx+1))
+
+/**
+ * cpupri_find - find the best (lowest-pri) CPU in the system
+ * @cp: The cpupri context
+ * @p: The task
+ * @lowest_mask: A mask to fill in with selected CPUs
+ *
+ * Note: This function returns the recommended CPUs as calculated during the
+ * current invokation. By the time the call returns, the CPUs may have in
+ * fact changed priorities any number of times. While not ideal, it is not
+ * an issue of correctness since the normal rebalancer logic will correct
+ * any discrepancies created by racing against the uncertainty of the current
+ * priority configuration.
+ *
+ * Returns: (int)bool - CPUs were found
+ */
+int cpupri_find(struct cpupri *cp, struct task_struct *p,
+ cpumask_t *lowest_mask)
+{
+ int idx = 0;
+ int task_pri = convert_prio(p->prio);
+
+ for_each_cpupri_active(cp->pri_active, idx) {
+ struct cpupri_vec *vec = &cp->pri_to_cpu[idx];
+ cpumask_t mask;
+
+ if (idx >= task_pri)
+ break;
+
+ cpus_and(mask, p->cpus_allowed, vec->mask);
+
+ if (cpus_empty(mask))
+ continue;
+
+ *lowest_mask = mask;
+ return 1;
+ }
+
+ return 0;
+}
+
+/**
+ * cpupri_set - update the cpu priority setting
+ * @cp: The cpupri context
+ * @cpu: The target cpu
+ * @pri: The priority (INVALID-RT99) to assign to this CPU
+ *
+ * Note: Assumes cpu_rq(cpu)->lock is locked
+ *
+ * Returns: (void)
+ */
+void cpupri_set(struct cpupri *cp, int cpu, int newpri)
+{
+ int *currpri = &cp->cpu_to_pri[cpu];
+ int oldpri = *currpri;
+ unsigned long flags;
+
+ newpri = convert_prio(newpri);
+
+ BUG_ON(newpri >= CPUPRI_NR_PRIORITIES);
+
+ if (newpri == oldpri)
+ return;
+
+ /*
+ * If the cpu was currently mapped to a different value, we
+ * first need to unmap the old value
+ */
+ if (likely(oldpri != CPUPRI_INVALID)) {
+ struct cpupri_vec *vec = &cp->pri_to_cpu[oldpri];
+
+ spin_lock_irqsave(&vec->lock, flags);
+
+ vec->count--;
+ if (!vec->count)
+ clear_bit(oldpri, cp->pri_active);
+ cpu_clear(cpu, vec->mask);
+
+ spin_unlock_irqrestore(&vec->lock, flags);
+ }
+
+ if (likely(newpri != CPUPRI_INVALID)) {
+ struct cpupri_vec *vec = &cp->pri_to_cpu[newpri];
+
+ spin_lock_irqsave(&vec->lock, flags);
+
+ cpu_set(cpu, vec->mask);
+ vec->count++;
+ if (vec->count == 1)
+ set_bit(newpri, cp->pri_active);
+
+ spin_unlock_irqrestore(&vec->lock, flags);
+ }
+
+ *currpri = newpri;
+}
+
+/**
+ * cpupri_init - initialize the cpupri structure
+ * @cp: The cpupri context
+ *
+ * Returns: (void)
+ */
+void cpupri_init(struct cpupri *cp)
+{
+ int i;
+
+ memset(cp, 0, sizeof(*cp));
+
+ for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) {
+ struct cpupri_vec *vec = &cp->pri_to_cpu[i];
+
+ spin_lock_init(&vec->lock);
+ vec->count = 0;
+ cpus_clear(vec->mask);
+ }
+
+ for_each_possible_cpu(i)
+ cp->cpu_to_pri[i] = CPUPRI_INVALID;
+}
+
+
diff --git a/kernel/sched_cpupri.h b/kernel/sched_cpupri.h
new file mode 100644
index 0000000..0b6a3d1
--- /dev/null
+++ b/kernel/sched_cpupri.h
@@ -0,0 +1,36 @@
+#ifndef _LINUX_CPUPRI_H
+#define _LINUX_CPUPRI_H
+
+#include <linux/sched.h>
+
+#define CPUPRI_NR_PRIORITIES 2+MAX_RT_PRIO
+#define CPUPRI_NR_PRI_WORDS CPUPRI_NR_PRIORITIES/BITS_PER_LONG
+
+#define CPUPRI_INVALID -1
+#define CPUPRI_IDLE 0
+#define CPUPRI_NORMAL 1
+/* values 2-101 are RT priorities 0-99 */
+
+struct cpupri_vec {
+ spinlock_t lock;
+ int count;
+ cpumask_t mask;
+};
+
+struct cpupri {
+ struct cpupri_vec pri_to_cpu[CPUPRI_NR_PRIORITIES];
+ long pri_active[CPUPRI_NR_PRI_WORDS];
+ int cpu_to_pri[NR_CPUS];
+};
+
+#ifdef CONFIG_SMP
+int cpupri_find(struct cpupri *cp,
+ struct task_struct *p, cpumask_t *lowest_mask);
+void cpupri_set(struct cpupri *cp, int cpu, int pri);
+void cpupri_init(struct cpupri *cp);
+#else
+#define cpupri_set(cp, cpu, pri) do { } while (0)
+#define cpupri_init() do { } while (0)
+#endif
+
+#endif /* _LINUX_CPUPRI_H */
diff --git a/kernel/sched_rt.c b/kernel/sched_rt.c
index 3cc98be..79e5ad5 100644
--- a/kernel/sched_rt.c
+++ b/kernel/sched_rt.c
@@ -391,8 +391,11 @@ void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
WARN_ON(!rt_prio(rt_se_prio(rt_se)));
rt_rq->rt_nr_running++;
#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
- if (rt_se_prio(rt_se) < rt_rq->highest_prio)
+ if (rt_se_prio(rt_se) < rt_rq->highest_prio) {
+ struct rq *rq = rq_of_rt_rq(rt_rq);
rt_rq->highest_prio = rt_se_prio(rt_se);
+ cpupri_set(&rq->rd->cpupri, rq->cpu, rt_se_prio(rt_se));
+ }
#endif
#ifdef CONFIG_SMP
if (rt_se->nr_cpus_allowed > 1) {
@@ -416,6 +419,10 @@ void inc_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
static inline
void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
{
+#ifdef CONFIG_SMP
+ int highest_prio = rt_rq->highest_prio;
+#endif
+
WARN_ON(!rt_prio(rt_se_prio(rt_se)));
WARN_ON(!rt_rq->rt_nr_running);
rt_rq->rt_nr_running--;
@@ -439,6 +446,11 @@ void dec_rt_tasks(struct sched_rt_entity *rt_se, struct rt_rq *rt_rq)
rq->rt.rt_nr_migratory--;
}

+ if (rt_rq->highest_prio != highest_prio) {
+ struct rq *rq = rq_of_rt_rq(rt_rq);
+ cpupri_set(&rq->rd->cpupri, rq->cpu, rt_rq->highest_prio);
+ }
+
update_rt_migration(rq_of_rt_rq(rt_rq));
#endif /* CONFIG_SMP */
#ifdef CONFIG_RT_GROUP_SCHED
@@ -723,73 +735,6 @@ static struct task_struct *pick_next_highest_task_rt(struct rq *rq, int cpu)

static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);

-static int find_lowest_cpus(struct task_struct *task, cpumask_t *lowest_mask)
-{
- int lowest_prio = -1;
- int lowest_cpu = -1;
- int count = 0;
- int cpu;
-
- cpus_and(*lowest_mask, task_rq(task)->rd->online, task->cpus_allowed);
-
- /*
- * Scan each rq for the lowest prio.
- */
- for_each_cpu_mask(cpu, *lowest_mask) {
- struct rq *rq = cpu_rq(cpu);
-
- /* We look for lowest RT prio or non-rt CPU */
- if (rq->rt.highest_prio >= MAX_RT_PRIO) {
- /*
- * if we already found a low RT queue
- * and now we found this non-rt queue
- * clear the mask and set our bit.
- * Otherwise just return the queue as is
- * and the count==1 will cause the algorithm
- * to use the first bit found.
- */
- if (lowest_cpu != -1) {
- cpus_clear(*lowest_mask);
- cpu_set(rq->cpu, *lowest_mask);
- }
- return 1;
- }
-
- /* no locking for now */
- if ((rq->rt.highest_prio > task->prio)
- && (rq->rt.highest_prio >= lowest_prio)) {
- if (rq->rt.highest_prio > lowest_prio) {
- /* new low - clear old data */
- lowest_prio = rq->rt.highest_prio;
- lowest_cpu = cpu;
- count = 0;
- }
- count++;
- } else
- cpu_clear(cpu, *lowest_mask);
- }
-
- /*
- * Clear out all the set bits that represent
- * runqueues that were of higher prio than
- * the lowest_prio.
- */
- if (lowest_cpu > 0) {
- /*
- * Perhaps we could add another cpumask op to
- * zero out bits. Like cpu_zero_bits(cpumask, nrbits);
- * Then that could be optimized to use memset and such.
- */
- for_each_cpu_mask(cpu, *lowest_mask) {
- if (cpu >= lowest_cpu)
- break;
- cpu_clear(cpu, *lowest_mask);
- }
- }
-
- return count;
-}
-
static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
{
int first;
@@ -811,17 +756,12 @@ static int find_lowest_rq(struct task_struct *task)
cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
int this_cpu = smp_processor_id();
int cpu = task_cpu(task);
- int count = find_lowest_cpus(task, lowest_mask);

- if (!count)
- return -1; /* No targets found */
+ if (task->rt.nr_cpus_allowed == 1)
+ return -1; /* No other targets possible */

- /*
- * There is no sense in performing an optimal search if only one
- * target is found.
- */
- if (count == 1)
- return first_cpu(*lowest_mask);
+ if (!cpupri_find(&task_rq(task)->rd->cpupri, task, lowest_mask))
+ return -1; /* No targets found */

/*
* At this point we have built a mask of cpus representing the
@@ -1178,6 +1118,8 @@ static void join_domain_rt(struct rq *rq)
{
if (rq->rt.overloaded)
rt_set_overload(rq);
+
+ cpupri_set(&rq->rd->cpupri, rq->cpu, rq->rt.highest_prio);
}

/* Assumes rq->lock is held */
@@ -1185,6 +1127,8 @@ static void leave_domain_rt(struct rq *rq)
{
if (rq->rt.overloaded)
rt_clear_overload(rq);
+
+ cpupri_set(&rq->rd->cpupri, rq->cpu, CPUPRI_INVALID);
}

/*

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/