sched.c

/* sched.c - SPU scheduler.
 *
 * Copyright (C) IBM 2005
 * Author: Mark Nutter <mnutter@us.ibm.com>
 *
 * 2006-03-31	NUMA domains added.
 *
 * 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; either version 2, or (at your option)
 * any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
 */

#undef DEBUG

#include <linux/module.h>
#include <linux/errno.h>
#include <linux/sched.h>
#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/completion.h>
#include <linux/vmalloc.h>
#include <linux/smp.h>
#include <linux/stddef.h>
#include <linux/unistd.h>
#include <linux/numa.h>
#include <linux/mutex.h>
#include <linux/notifier.h>
#include <linux/kthread.h>
#include <linux/pid_namespace.h>
#include <linux/proc_fs.h>
#include <linux/seq_file.h>
#include <linux/marker.h>

#include <asm/io.h>
#include <asm/mmu_context.h>
#include <asm/spu.h>
#include <asm/spu_csa.h>
#include <asm/spu_priv1.h>
#include "spufs.h"

struct spu_prio_array {
	DECLARE_BITMAP(bitmap, MAX_PRIO);
	struct list_head runq[MAX_PRIO];
	spinlock_t runq_lock;
	int nr_waiting;
};

static unsigned long spu_avenrun[3];
static struct spu_prio_array *spu_prio;
static struct task_struct *spusched_task;
static struct timer_list spusched_timer;
static struct timer_list spuloadavg_timer;

/*
 * Priority of a normal, non-rt, non-niced'd process (aka nice level 0).
 */
#define NORMAL_PRIO		120

/*
 * Frequency of the spu scheduler tick.  By default we do one SPU scheduler
 * tick for every 10 CPU scheduler ticks.
 */
#define SPUSCHED_TICK		(10)

/*
 * These are the 'tuning knobs' of the scheduler:
 *
 * Minimum timeslice is 5 msecs (or 1 spu scheduler tick, whichever is
 * larger), default timeslice is 100 msecs, maximum timeslice is 800 msecs.
 */
#define MIN_SPU_TIMESLICE	max(5 * HZ / (1000 * SPUSCHED_TICK), 1)
#define DEF_SPU_TIMESLICE	(100 * HZ / (1000 * SPUSCHED_TICK))

#define MAX_USER_PRIO		(MAX_PRIO - MAX_RT_PRIO)
#define SCALE_PRIO(x, prio) \
	max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_SPU_TIMESLICE)

/*
 * scale user-nice values [ -20 ... 0 ... 19 ] to time slice values:
 * [800ms ... 100ms ... 5ms]
 *
 * The higher a thread's priority, the bigger timeslices
 * it gets during one round of execution. But even the lowest
 * priority thread gets MIN_TIMESLICE worth of execution time.
 */
void spu_set_timeslice(struct spu_context *ctx)
{
	if (ctx->prio < NORMAL_PRIO)
		ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE * 4, ctx->prio);
	else
		ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE, ctx->prio);
}

/*
 * Update scheduling information from the owning thread.
 */
void __spu_update_sched_info(struct spu_context *ctx)
{
	/*
	 * assert that the context is not on the runqueue, so it is safe
	 * to change its scheduling parameters.
	 */
	BUG_ON(!list_empty(&ctx->rq));

	/*
	 * 32-Bit assignments are atomic on powerpc, and we don't care about
	 * memory ordering here because retrieving the controlling thread is
	 * per definition racy.
	 */
	ctx->tid = current->pid;

	/*
	 * We do our own priority calculations, so we normally want
	 * ->static_prio to start with. Unfortunately this field
	 * contains junk for threads with a realtime scheduling
	 * policy so we have to look at ->prio in this case.
	 */
	if (rt_prio(current->prio))
		ctx->prio = current->prio;
	else
		ctx->prio = current->static_prio;
	ctx->policy = current->policy;

	/*
	 * TO DO: the context may be loaded, so we may need to activate
	 * it again on a different node. But it shouldn't hurt anything
	 * to update its parameters, because we know that the scheduler
	 * is not actively looking at this field, since it is not on the
	 * runqueue. The context will be rescheduled on the proper node
	 * if it is timesliced or preempted.
	 */
	ctx->cpus_allowed = current->cpus_allowed;

	/* Save the current cpu id for spu interrupt routing. */
	ctx->last_ran = raw_smp_processor_id();
}

void spu_update_sched_info(struct spu_context *ctx)
{
	int node;

	if (ctx->state == SPU_STATE_RUNNABLE) {
		node = ctx->spu->node;

		/*
		 * Take list_mutex to sync with find_victim().
		 */
		mutex_lock(&cbe_spu_info[node].list_mutex);
		__spu_update_sched_info(ctx);
		mutex_unlock(&cbe_spu_info[node].list_mutex);
	} else {
		__spu_update_sched_info(ctx);
	}
}

static int __node_allowed(struct spu_context *ctx, int node)
{
	if (nr_cpus_node(node)) {
		cpumask_t mask = node_to_cpumask(node);

		if (cpus_intersects(mask, ctx->cpus_allowed))
			return 1;
	}

	return 0;
}

static int node_allowed(struct spu_context *ctx, int node)
{
	int rval;

	spin_lock(&spu_prio->runq_lock);
	rval = __node_allowed(ctx, node);
	spin_unlock(&spu_prio->runq_lock);

	return rval;
}

void do_notify_spus_active(void)
{
	int node;

	/*
	 * Wake up the active spu_contexts.
	 *
	 * When the awakened processes see their "notify_active" flag is set,
	 * they will call spu_switch_notify().
	 */
	for_each_online_node(node) {
		struct spu *spu;