PageRenderTime 110ms CodeModel.GetById 30ms app.highlight 71ms RepoModel.GetById 1ms app.codeStats 0ms

/arch/powerpc/platforms/cell/spufs/sched.c

http://github.com/mirrors/linux
C | 1145 lines | 745 code | 168 blank | 232 comment | 126 complexity | cc91f99f5bb5e347c457ea5ceae8b53f MD5 | raw file
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/* sched.c - SPU scheduler.
   3 *
   4 * Copyright (C) IBM 2005
   5 * Author: Mark Nutter <mnutter@us.ibm.com>
   6 *
   7 * 2006-03-31	NUMA domains added.
   8 */
   9
  10#undef DEBUG
  11
  12#include <linux/errno.h>
  13#include <linux/sched/signal.h>
  14#include <linux/sched/loadavg.h>
  15#include <linux/sched/rt.h>
  16#include <linux/kernel.h>
  17#include <linux/mm.h>
  18#include <linux/slab.h>
  19#include <linux/completion.h>
  20#include <linux/vmalloc.h>
  21#include <linux/smp.h>
  22#include <linux/stddef.h>
  23#include <linux/unistd.h>
  24#include <linux/numa.h>
  25#include <linux/mutex.h>
  26#include <linux/notifier.h>
  27#include <linux/kthread.h>
  28#include <linux/pid_namespace.h>
  29#include <linux/proc_fs.h>
  30#include <linux/seq_file.h>
  31
  32#include <asm/io.h>
  33#include <asm/mmu_context.h>
  34#include <asm/spu.h>
  35#include <asm/spu_csa.h>
  36#include <asm/spu_priv1.h>
  37#include "spufs.h"
  38#define CREATE_TRACE_POINTS
  39#include "sputrace.h"
  40
  41struct spu_prio_array {
  42	DECLARE_BITMAP(bitmap, MAX_PRIO);
  43	struct list_head runq[MAX_PRIO];
  44	spinlock_t runq_lock;
  45	int nr_waiting;
  46};
  47
  48static unsigned long spu_avenrun[3];
  49static struct spu_prio_array *spu_prio;
  50static struct task_struct *spusched_task;
  51static struct timer_list spusched_timer;
  52static struct timer_list spuloadavg_timer;
  53
  54/*
  55 * Priority of a normal, non-rt, non-niced'd process (aka nice level 0).
  56 */
  57#define NORMAL_PRIO		120
  58
  59/*
  60 * Frequency of the spu scheduler tick.  By default we do one SPU scheduler
  61 * tick for every 10 CPU scheduler ticks.
  62 */
  63#define SPUSCHED_TICK		(10)
  64
  65/*
  66 * These are the 'tuning knobs' of the scheduler:
  67 *
  68 * Minimum timeslice is 5 msecs (or 1 spu scheduler tick, whichever is
  69 * larger), default timeslice is 100 msecs, maximum timeslice is 800 msecs.
  70 */
  71#define MIN_SPU_TIMESLICE	max(5 * HZ / (1000 * SPUSCHED_TICK), 1)
  72#define DEF_SPU_TIMESLICE	(100 * HZ / (1000 * SPUSCHED_TICK))
  73
  74#define SCALE_PRIO(x, prio) \
  75	max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_SPU_TIMESLICE)
  76
  77/*
  78 * scale user-nice values [ -20 ... 0 ... 19 ] to time slice values:
  79 * [800ms ... 100ms ... 5ms]
  80 *
  81 * The higher a thread's priority, the bigger timeslices
  82 * it gets during one round of execution. But even the lowest
  83 * priority thread gets MIN_TIMESLICE worth of execution time.
  84 */
  85void spu_set_timeslice(struct spu_context *ctx)
  86{
  87	if (ctx->prio < NORMAL_PRIO)
  88		ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE * 4, ctx->prio);
  89	else
  90		ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE, ctx->prio);
  91}
  92
  93/*
  94 * Update scheduling information from the owning thread.
  95 */
  96void __spu_update_sched_info(struct spu_context *ctx)
  97{
  98	/*
  99	 * assert that the context is not on the runqueue, so it is safe
 100	 * to change its scheduling parameters.
 101	 */
 102	BUG_ON(!list_empty(&ctx->rq));
 103
 104	/*
 105	 * 32-Bit assignments are atomic on powerpc, and we don't care about
 106	 * memory ordering here because retrieving the controlling thread is
 107	 * per definition racy.
 108	 */
 109	ctx->tid = current->pid;
 110
 111	/*
 112	 * We do our own priority calculations, so we normally want
 113	 * ->static_prio to start with. Unfortunately this field
 114	 * contains junk for threads with a realtime scheduling
 115	 * policy so we have to look at ->prio in this case.
 116	 */
 117	if (rt_prio(current->prio))
 118		ctx->prio = current->prio;
 119	else
 120		ctx->prio = current->static_prio;
 121	ctx->policy = current->policy;
 122
 123	/*
 124	 * TO DO: the context may be loaded, so we may need to activate
 125	 * it again on a different node. But it shouldn't hurt anything
 126	 * to update its parameters, because we know that the scheduler
 127	 * is not actively looking at this field, since it is not on the
 128	 * runqueue. The context will be rescheduled on the proper node
 129	 * if it is timesliced or preempted.
 130	 */
 131	cpumask_copy(&ctx->cpus_allowed, current->cpus_ptr);
 132
 133	/* Save the current cpu id for spu interrupt routing. */
 134	ctx->last_ran = raw_smp_processor_id();
 135}
 136
 137void spu_update_sched_info(struct spu_context *ctx)
 138{
 139	int node;
 140
 141	if (ctx->state == SPU_STATE_RUNNABLE) {
 142		node = ctx->spu->node;
 143
 144		/*
 145		 * Take list_mutex to sync with find_victim().
 146		 */
 147		mutex_lock(&cbe_spu_info[node].list_mutex);
 148		__spu_update_sched_info(ctx);
 149		mutex_unlock(&cbe_spu_info[node].list_mutex);
 150	} else {
 151		__spu_update_sched_info(ctx);
 152	}
 153}
 154
 155static int __node_allowed(struct spu_context *ctx, int node)
 156{
 157	if (nr_cpus_node(node)) {
 158		const struct cpumask *mask = cpumask_of_node(node);
 159
 160		if (cpumask_intersects(mask, &ctx->cpus_allowed))
 161			return 1;
 162	}
 163
 164	return 0;
 165}
 166
 167static int node_allowed(struct spu_context *ctx, int node)
 168{
 169	int rval;
 170
 171	spin_lock(&spu_prio->runq_lock);
 172	rval = __node_allowed(ctx, node);
 173	spin_unlock(&spu_prio->runq_lock);
 174
 175	return rval;
 176}
 177
 178void do_notify_spus_active(void)
 179{
 180	int node;
 181
 182	/*
 183	 * Wake up the active spu_contexts.
 184	 *
 185	 * When the awakened processes see their "notify_active" flag is set,
 186	 * they will call spu_switch_notify().
 187	 */
 188	for_each_online_node(node) {
 189		struct spu *spu;
 190
 191		mutex_lock(&cbe_spu_info[node].list_mutex);
 192		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
 193			if (spu->alloc_state != SPU_FREE) {
 194				struct spu_context *ctx = spu->ctx;
 195				set_bit(SPU_SCHED_NOTIFY_ACTIVE,
 196					&ctx->sched_flags);
 197				mb();
 198				wake_up_all(&ctx->stop_wq);
 199			}
 200		}
 201		mutex_unlock(&cbe_spu_info[node].list_mutex);
 202	}
 203}
 204
 205/**
 206 * spu_bind_context - bind spu context to physical spu
 207 * @spu:	physical spu to bind to
 208 * @ctx:	context to bind
 209 */
 210static void spu_bind_context(struct spu *spu, struct spu_context *ctx)
 211{
 212	spu_context_trace(spu_bind_context__enter, ctx, spu);
 213
 214	spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
 215
 216	if (ctx->flags & SPU_CREATE_NOSCHED)
 217		atomic_inc(&cbe_spu_info[spu->node].reserved_spus);
 218
 219	ctx->stats.slb_flt_base = spu->stats.slb_flt;
 220	ctx->stats.class2_intr_base = spu->stats.class2_intr;
 221
 222	spu_associate_mm(spu, ctx->owner);
 223
 224	spin_lock_irq(&spu->register_lock);
 225	spu->ctx = ctx;
 226	spu->flags = 0;
 227	ctx->spu = spu;
 228	ctx->ops = &spu_hw_ops;
 229	spu->pid = current->pid;
 230	spu->tgid = current->tgid;
 231	spu->ibox_callback = spufs_ibox_callback;
 232	spu->wbox_callback = spufs_wbox_callback;
 233	spu->stop_callback = spufs_stop_callback;
 234	spu->mfc_callback = spufs_mfc_callback;
 235	spin_unlock_irq(&spu->register_lock);
 236
 237	spu_unmap_mappings(ctx);
 238
 239	spu_switch_log_notify(spu, ctx, SWITCH_LOG_START, 0);
 240	spu_restore(&ctx->csa, spu);
 241	spu->timestamp = jiffies;
 242	spu_switch_notify(spu, ctx);
 243	ctx->state = SPU_STATE_RUNNABLE;
 244
 245	spuctx_switch_state(ctx, SPU_UTIL_USER);
 246}
 247
 248/*
 249 * Must be used with the list_mutex held.
 250 */
 251static inline int sched_spu(struct spu *spu)
 252{
 253	BUG_ON(!mutex_is_locked(&cbe_spu_info[spu->node].list_mutex));
 254
 255	return (!spu->ctx || !(spu->ctx->flags & SPU_CREATE_NOSCHED));
 256}
 257
 258static void aff_merge_remaining_ctxs(struct spu_gang *gang)
 259{
 260	struct spu_context *ctx;
 261
 262	list_for_each_entry(ctx, &gang->aff_list_head, aff_list) {
 263		if (list_empty(&ctx->aff_list))
 264			list_add(&ctx->aff_list, &gang->aff_list_head);
 265	}
 266	gang->aff_flags |= AFF_MERGED;
 267}
 268
 269static void aff_set_offsets(struct spu_gang *gang)
 270{
 271	struct spu_context *ctx;
 272	int offset;
 273
 274	offset = -1;
 275	list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
 276								aff_list) {
 277		if (&ctx->aff_list == &gang->aff_list_head)
 278			break;
 279		ctx->aff_offset = offset--;
 280	}
 281
 282	offset = 0;
 283	list_for_each_entry(ctx, gang->aff_ref_ctx->aff_list.prev, aff_list) {
 284		if (&ctx->aff_list == &gang->aff_list_head)
 285			break;
 286		ctx->aff_offset = offset++;
 287	}
 288
 289	gang->aff_flags |= AFF_OFFSETS_SET;
 290}
 291
 292static struct spu *aff_ref_location(struct spu_context *ctx, int mem_aff,
 293		 int group_size, int lowest_offset)
 294{
 295	struct spu *spu;
 296	int node, n;
 297
 298	/*
 299	 * TODO: A better algorithm could be used to find a good spu to be
 300	 *       used as reference location for the ctxs chain.
 301	 */
 302	node = cpu_to_node(raw_smp_processor_id());
 303	for (n = 0; n < MAX_NUMNODES; n++, node++) {
 304		/*
 305		 * "available_spus" counts how many spus are not potentially
 306		 * going to be used by other affinity gangs whose reference
 307		 * context is already in place. Although this code seeks to
 308		 * avoid having affinity gangs with a summed amount of
 309		 * contexts bigger than the amount of spus in the node,
 310		 * this may happen sporadically. In this case, available_spus
 311		 * becomes negative, which is harmless.
 312		 */
 313		int available_spus;
 314
 315		node = (node < MAX_NUMNODES) ? node : 0;
 316		if (!node_allowed(ctx, node))
 317			continue;
 318
 319		available_spus = 0;
 320		mutex_lock(&cbe_spu_info[node].list_mutex);
 321		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
 322			if (spu->ctx && spu->ctx->gang && !spu->ctx->aff_offset
 323					&& spu->ctx->gang->aff_ref_spu)
 324				available_spus -= spu->ctx->gang->contexts;
 325			available_spus++;
 326		}
 327		if (available_spus < ctx->gang->contexts) {
 328			mutex_unlock(&cbe_spu_info[node].list_mutex);
 329			continue;
 330		}
 331
 332		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
 333			if ((!mem_aff || spu->has_mem_affinity) &&
 334							sched_spu(spu)) {
 335				mutex_unlock(&cbe_spu_info[node].list_mutex);
 336				return spu;
 337			}
 338		}
 339		mutex_unlock(&cbe_spu_info[node].list_mutex);
 340	}
 341	return NULL;
 342}
 343
 344static void aff_set_ref_point_location(struct spu_gang *gang)
 345{
 346	int mem_aff, gs, lowest_offset;
 347	struct spu_context *ctx;
 348	struct spu *tmp;
 349
 350	mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM;
 351	lowest_offset = 0;
 352	gs = 0;
 353
 354	list_for_each_entry(tmp, &gang->aff_list_head, aff_list)
 355		gs++;
 356
 357	list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
 358								aff_list) {
 359		if (&ctx->aff_list == &gang->aff_list_head)
 360			break;
 361		lowest_offset = ctx->aff_offset;
 362	}
 363
 364	gang->aff_ref_spu = aff_ref_location(gang->aff_ref_ctx, mem_aff, gs,
 365							lowest_offset);
 366}
 367
 368static struct spu *ctx_location(struct spu *ref, int offset, int node)
 369{
 370	struct spu *spu;
 371
 372	spu = NULL;
 373	if (offset >= 0) {
 374		list_for_each_entry(spu, ref->aff_list.prev, aff_list) {
 375			BUG_ON(spu->node != node);
 376			if (offset == 0)
 377				break;
 378			if (sched_spu(spu))
 379				offset--;
 380		}
 381	} else {
 382		list_for_each_entry_reverse(spu, ref->aff_list.next, aff_list) {
 383			BUG_ON(spu->node != node);
 384			if (offset == 0)
 385				break;
 386			if (sched_spu(spu))
 387				offset++;
 388		}
 389	}
 390
 391	return spu;
 392}
 393
 394/*
 395 * affinity_check is called each time a context is going to be scheduled.
 396 * It returns the spu ptr on which the context must run.
 397 */
 398static int has_affinity(struct spu_context *ctx)
 399{
 400	struct spu_gang *gang = ctx->gang;
 401
 402	if (list_empty(&ctx->aff_list))
 403		return 0;
 404
 405	if (atomic_read(&ctx->gang->aff_sched_count) == 0)
 406		ctx->gang->aff_ref_spu = NULL;
 407
 408	if (!gang->aff_ref_spu) {
 409		if (!(gang->aff_flags & AFF_MERGED))
 410			aff_merge_remaining_ctxs(gang);
 411		if (!(gang->aff_flags & AFF_OFFSETS_SET))
 412			aff_set_offsets(gang);
 413		aff_set_ref_point_location(gang);
 414	}
 415
 416	return gang->aff_ref_spu != NULL;
 417}
 418
 419/**
 420 * spu_unbind_context - unbind spu context from physical spu
 421 * @spu:	physical spu to unbind from
 422 * @ctx:	context to unbind
 423 */
 424static void spu_unbind_context(struct spu *spu, struct spu_context *ctx)
 425{
 426	u32 status;
 427
 428	spu_context_trace(spu_unbind_context__enter, ctx, spu);
 429
 430	spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
 431
 432 	if (spu->ctx->flags & SPU_CREATE_NOSCHED)
 433		atomic_dec(&cbe_spu_info[spu->node].reserved_spus);
 434
 435	if (ctx->gang)
 436		/*
 437		 * If ctx->gang->aff_sched_count is positive, SPU affinity is
 438		 * being considered in this gang. Using atomic_dec_if_positive
 439		 * allow us to skip an explicit check for affinity in this gang
 440		 */
 441		atomic_dec_if_positive(&ctx->gang->aff_sched_count);
 442
 443	spu_switch_notify(spu, NULL);
 444	spu_unmap_mappings(ctx);
 445	spu_save(&ctx->csa, spu);
 446	spu_switch_log_notify(spu, ctx, SWITCH_LOG_STOP, 0);
 447
 448	spin_lock_irq(&spu->register_lock);
 449	spu->timestamp = jiffies;
 450	ctx->state = SPU_STATE_SAVED;
 451	spu->ibox_callback = NULL;
 452	spu->wbox_callback = NULL;
 453	spu->stop_callback = NULL;
 454	spu->mfc_callback = NULL;
 455	spu->pid = 0;
 456	spu->tgid = 0;
 457	ctx->ops = &spu_backing_ops;
 458	spu->flags = 0;
 459	spu->ctx = NULL;
 460	spin_unlock_irq(&spu->register_lock);
 461
 462	spu_associate_mm(spu, NULL);
 463
 464	ctx->stats.slb_flt +=
 465		(spu->stats.slb_flt - ctx->stats.slb_flt_base);
 466	ctx->stats.class2_intr +=
 467		(spu->stats.class2_intr - ctx->stats.class2_intr_base);
 468
 469	/* This maps the underlying spu state to idle */
 470	spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED);
 471	ctx->spu = NULL;
 472
 473	if (spu_stopped(ctx, &status))
 474		wake_up_all(&ctx->stop_wq);
 475}
 476
 477/**
 478 * spu_add_to_rq - add a context to the runqueue
 479 * @ctx:       context to add
 480 */
 481static void __spu_add_to_rq(struct spu_context *ctx)
 482{
 483	/*
 484	 * Unfortunately this code path can be called from multiple threads
 485	 * on behalf of a single context due to the way the problem state
 486	 * mmap support works.
 487	 *
 488	 * Fortunately we need to wake up all these threads at the same time
 489	 * and can simply skip the runqueue addition for every but the first
 490	 * thread getting into this codepath.
 491	 *
 492	 * It's still quite hacky, and long-term we should proxy all other
 493	 * threads through the owner thread so that spu_run is in control
 494	 * of all the scheduling activity for a given context.
 495	 */
 496	if (list_empty(&ctx->rq)) {
 497		list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]);
 498		set_bit(ctx->prio, spu_prio->bitmap);
 499		if (!spu_prio->nr_waiting++)
 500			mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
 501	}
 502}
 503
 504static void spu_add_to_rq(struct spu_context *ctx)
 505{
 506	spin_lock(&spu_prio->runq_lock);
 507	__spu_add_to_rq(ctx);
 508	spin_unlock(&spu_prio->runq_lock);
 509}
 510
 511static void __spu_del_from_rq(struct spu_context *ctx)
 512{
 513	int prio = ctx->prio;
 514
 515	if (!list_empty(&ctx->rq)) {
 516		if (!--spu_prio->nr_waiting)
 517			del_timer(&spusched_timer);
 518		list_del_init(&ctx->rq);
 519
 520		if (list_empty(&spu_prio->runq[prio]))
 521			clear_bit(prio, spu_prio->bitmap);
 522	}
 523}
 524
 525void spu_del_from_rq(struct spu_context *ctx)
 526{
 527	spin_lock(&spu_prio->runq_lock);
 528	__spu_del_from_rq(ctx);
 529	spin_unlock(&spu_prio->runq_lock);
 530}
 531
 532static void spu_prio_wait(struct spu_context *ctx)
 533{
 534	DEFINE_WAIT(wait);
 535
 536	/*
 537	 * The caller must explicitly wait for a context to be loaded
 538	 * if the nosched flag is set.  If NOSCHED is not set, the caller
 539	 * queues the context and waits for an spu event or error.
 540	 */
 541	BUG_ON(!(ctx->flags & SPU_CREATE_NOSCHED));
 542
 543	spin_lock(&spu_prio->runq_lock);
 544	prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE);
 545	if (!signal_pending(current)) {
 546		__spu_add_to_rq(ctx);
 547		spin_unlock(&spu_prio->runq_lock);
 548		mutex_unlock(&ctx->state_mutex);
 549		schedule();
 550		mutex_lock(&ctx->state_mutex);
 551		spin_lock(&spu_prio->runq_lock);
 552		__spu_del_from_rq(ctx);
 553	}
 554	spin_unlock(&spu_prio->runq_lock);
 555	__set_current_state(TASK_RUNNING);
 556	remove_wait_queue(&ctx->stop_wq, &wait);
 557}
 558
 559static struct spu *spu_get_idle(struct spu_context *ctx)
 560{
 561	struct spu *spu, *aff_ref_spu;
 562	int node, n;
 563
 564	spu_context_nospu_trace(spu_get_idle__enter, ctx);
 565
 566	if (ctx->gang) {
 567		mutex_lock(&ctx->gang->aff_mutex);
 568		if (has_affinity(ctx)) {
 569			aff_ref_spu = ctx->gang->aff_ref_spu;
 570			atomic_inc(&ctx->gang->aff_sched_count);
 571			mutex_unlock(&ctx->gang->aff_mutex);
 572			node = aff_ref_spu->node;
 573
 574			mutex_lock(&cbe_spu_info[node].list_mutex);
 575			spu = ctx_location(aff_ref_spu, ctx->aff_offset, node);
 576			if (spu && spu->alloc_state == SPU_FREE)
 577				goto found;
 578			mutex_unlock(&cbe_spu_info[node].list_mutex);
 579
 580			atomic_dec(&ctx->gang->aff_sched_count);
 581			goto not_found;
 582		}
 583		mutex_unlock(&ctx->gang->aff_mutex);
 584	}
 585	node = cpu_to_node(raw_smp_processor_id());
 586	for (n = 0; n < MAX_NUMNODES; n++, node++) {
 587		node = (node < MAX_NUMNODES) ? node : 0;
 588		if (!node_allowed(ctx, node))
 589			continue;
 590
 591		mutex_lock(&cbe_spu_info[node].list_mutex);
 592		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
 593			if (spu->alloc_state == SPU_FREE)
 594				goto found;
 595		}
 596		mutex_unlock(&cbe_spu_info[node].list_mutex);
 597	}
 598
 599 not_found:
 600	spu_context_nospu_trace(spu_get_idle__not_found, ctx);
 601	return NULL;
 602
 603 found:
 604	spu->alloc_state = SPU_USED;
 605	mutex_unlock(&cbe_spu_info[node].list_mutex);
 606	spu_context_trace(spu_get_idle__found, ctx, spu);
 607	spu_init_channels(spu);
 608	return spu;
 609}
 610
 611/**
 612 * find_victim - find a lower priority context to preempt
 613 * @ctx:	candidate context for running
 614 *
 615 * Returns the freed physical spu to run the new context on.
 616 */
 617static struct spu *find_victim(struct spu_context *ctx)
 618{
 619	struct spu_context *victim = NULL;
 620	struct spu *spu;
 621	int node, n;
 622
 623	spu_context_nospu_trace(spu_find_victim__enter, ctx);
 624
 625	/*
 626	 * Look for a possible preemption candidate on the local node first.
 627	 * If there is no candidate look at the other nodes.  This isn't
 628	 * exactly fair, but so far the whole spu scheduler tries to keep
 629	 * a strong node affinity.  We might want to fine-tune this in
 630	 * the future.
 631	 */
 632 restart:
 633	node = cpu_to_node(raw_smp_processor_id());
 634	for (n = 0; n < MAX_NUMNODES; n++, node++) {
 635		node = (node < MAX_NUMNODES) ? node : 0;
 636		if (!node_allowed(ctx, node))
 637			continue;
 638
 639		mutex_lock(&cbe_spu_info[node].list_mutex);
 640		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
 641			struct spu_context *tmp = spu->ctx;
 642
 643			if (tmp && tmp->prio > ctx->prio &&
 644			    !(tmp->flags & SPU_CREATE_NOSCHED) &&
 645			    (!victim || tmp->prio > victim->prio)) {
 646				victim = spu->ctx;
 647			}
 648		}
 649		if (victim)
 650			get_spu_context(victim);
 651		mutex_unlock(&cbe_spu_info[node].list_mutex);
 652
 653		if (victim) {
 654			/*
 655			 * This nests ctx->state_mutex, but we always lock
 656			 * higher priority contexts before lower priority
 657			 * ones, so this is safe until we introduce
 658			 * priority inheritance schemes.
 659			 *
 660			 * XXX if the highest priority context is locked,
 661			 * this can loop a long time.  Might be better to
 662			 * look at another context or give up after X retries.
 663			 */
 664			if (!mutex_trylock(&victim->state_mutex)) {
 665				put_spu_context(victim);
 666				victim = NULL;
 667				goto restart;
 668			}
 669
 670			spu = victim->spu;
 671			if (!spu || victim->prio <= ctx->prio) {
 672				/*
 673				 * This race can happen because we've dropped
 674				 * the active list mutex.  Not a problem, just
 675				 * restart the search.
 676				 */
 677				mutex_unlock(&victim->state_mutex);
 678				put_spu_context(victim);
 679				victim = NULL;
 680				goto restart;
 681			}
 682
 683			spu_context_trace(__spu_deactivate__unload, ctx, spu);
 684
 685			mutex_lock(&cbe_spu_info[node].list_mutex);
 686			cbe_spu_info[node].nr_active--;
 687			spu_unbind_context(spu, victim);
 688			mutex_unlock(&cbe_spu_info[node].list_mutex);
 689
 690			victim->stats.invol_ctx_switch++;
 691			spu->stats.invol_ctx_switch++;
 692			if (test_bit(SPU_SCHED_SPU_RUN, &victim->sched_flags))
 693				spu_add_to_rq(victim);
 694
 695			mutex_unlock(&victim->state_mutex);
 696			put_spu_context(victim);
 697
 698			return spu;
 699		}
 700	}
 701
 702	return NULL;
 703}
 704
 705static void __spu_schedule(struct spu *spu, struct spu_context *ctx)
 706{
 707	int node = spu->node;
 708	int success = 0;
 709
 710	spu_set_timeslice(ctx);
 711
 712	mutex_lock(&cbe_spu_info[node].list_mutex);
 713	if (spu->ctx == NULL) {
 714		spu_bind_context(spu, ctx);
 715		cbe_spu_info[node].nr_active++;
 716		spu->alloc_state = SPU_USED;
 717		success = 1;
 718	}
 719	mutex_unlock(&cbe_spu_info[node].list_mutex);
 720
 721	if (success)
 722		wake_up_all(&ctx->run_wq);
 723	else
 724		spu_add_to_rq(ctx);
 725}
 726
 727static void spu_schedule(struct spu *spu, struct spu_context *ctx)
 728{
 729	/* not a candidate for interruptible because it's called either
 730	   from the scheduler thread or from spu_deactivate */
 731	mutex_lock(&ctx->state_mutex);
 732	if (ctx->state == SPU_STATE_SAVED)
 733		__spu_schedule(spu, ctx);
 734	spu_release(ctx);
 735}
 736
 737/**
 738 * spu_unschedule - remove a context from a spu, and possibly release it.
 739 * @spu:	The SPU to unschedule from
 740 * @ctx:	The context currently scheduled on the SPU
 741 * @free_spu	Whether to free the SPU for other contexts
 742 *
 743 * Unbinds the context @ctx from the SPU @spu. If @free_spu is non-zero, the
 744 * SPU is made available for other contexts (ie, may be returned by
 745 * spu_get_idle). If this is zero, the caller is expected to schedule another
 746 * context to this spu.
 747 *
 748 * Should be called with ctx->state_mutex held.
 749 */
 750static void spu_unschedule(struct spu *spu, struct spu_context *ctx,
 751		int free_spu)
 752{
 753	int node = spu->node;
 754
 755	mutex_lock(&cbe_spu_info[node].list_mutex);
 756	cbe_spu_info[node].nr_active--;
 757	if (free_spu)
 758		spu->alloc_state = SPU_FREE;
 759	spu_unbind_context(spu, ctx);
 760	ctx->stats.invol_ctx_switch++;
 761	spu->stats.invol_ctx_switch++;
 762	mutex_unlock(&cbe_spu_info[node].list_mutex);
 763}
 764
 765/**
 766 * spu_activate - find a free spu for a context and execute it
 767 * @ctx:	spu context to schedule
 768 * @flags:	flags (currently ignored)
 769 *
 770 * Tries to find a free spu to run @ctx.  If no free spu is available
 771 * add the context to the runqueue so it gets woken up once an spu
 772 * is available.
 773 */
 774int spu_activate(struct spu_context *ctx, unsigned long flags)
 775{
 776	struct spu *spu;
 777
 778	/*
 779	 * If there are multiple threads waiting for a single context
 780	 * only one actually binds the context while the others will
 781	 * only be able to acquire the state_mutex once the context
 782	 * already is in runnable state.
 783	 */
 784	if (ctx->spu)
 785		return 0;
 786
 787spu_activate_top:
 788	if (signal_pending(current))
 789		return -ERESTARTSYS;
 790
 791	spu = spu_get_idle(ctx);
 792	/*
 793	 * If this is a realtime thread we try to get it running by
 794	 * preempting a lower priority thread.
 795	 */
 796	if (!spu && rt_prio(ctx->prio))
 797		spu = find_victim(ctx);
 798	if (spu) {
 799		unsigned long runcntl;
 800
 801		runcntl = ctx->ops->runcntl_read(ctx);
 802		__spu_schedule(spu, ctx);
 803		if (runcntl & SPU_RUNCNTL_RUNNABLE)
 804			spuctx_switch_state(ctx, SPU_UTIL_USER);
 805
 806		return 0;
 807	}
 808
 809	if (ctx->flags & SPU_CREATE_NOSCHED) {
 810		spu_prio_wait(ctx);
 811		goto spu_activate_top;
 812	}
 813
 814	spu_add_to_rq(ctx);
 815
 816	return 0;
 817}
 818
 819/**
 820 * grab_runnable_context - try to find a runnable context
 821 *
 822 * Remove the highest priority context on the runqueue and return it
 823 * to the caller.  Returns %NULL if no runnable context was found.
 824 */
 825static struct spu_context *grab_runnable_context(int prio, int node)
 826{
 827	struct spu_context *ctx;
 828	int best;
 829
 830	spin_lock(&spu_prio->runq_lock);
 831	best = find_first_bit(spu_prio->bitmap, prio);
 832	while (best < prio) {
 833		struct list_head *rq = &spu_prio->runq[best];
 834
 835		list_for_each_entry(ctx, rq, rq) {
 836			/* XXX(hch): check for affinity here as well */
 837			if (__node_allowed(ctx, node)) {
 838				__spu_del_from_rq(ctx);
 839				goto found;
 840			}
 841		}
 842		best++;
 843	}
 844	ctx = NULL;
 845 found:
 846	spin_unlock(&spu_prio->runq_lock);
 847	return ctx;
 848}
 849
 850static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio)
 851{
 852	struct spu *spu = ctx->spu;
 853	struct spu_context *new = NULL;
 854
 855	if (spu) {
 856		new = grab_runnable_context(max_prio, spu->node);
 857		if (new || force) {
 858			spu_unschedule(spu, ctx, new == NULL);
 859			if (new) {
 860				if (new->flags & SPU_CREATE_NOSCHED)
 861					wake_up(&new->stop_wq);
 862				else {
 863					spu_release(ctx);
 864					spu_schedule(spu, new);
 865					/* this one can't easily be made
 866					   interruptible */
 867					mutex_lock(&ctx->state_mutex);
 868				}
 869			}
 870		}
 871	}
 872
 873	return new != NULL;
 874}
 875
 876/**
 877 * spu_deactivate - unbind a context from it's physical spu
 878 * @ctx:	spu context to unbind
 879 *
 880 * Unbind @ctx from the physical spu it is running on and schedule
 881 * the highest priority context to run on the freed physical spu.
 882 */
 883void spu_deactivate(struct spu_context *ctx)
 884{
 885	spu_context_nospu_trace(spu_deactivate__enter, ctx);
 886	__spu_deactivate(ctx, 1, MAX_PRIO);
 887}
 888
 889/**
 890 * spu_yield -	yield a physical spu if others are waiting
 891 * @ctx:	spu context to yield
 892 *
 893 * Check if there is a higher priority context waiting and if yes
 894 * unbind @ctx from the physical spu and schedule the highest
 895 * priority context to run on the freed physical spu instead.
 896 */
 897void spu_yield(struct spu_context *ctx)
 898{
 899	spu_context_nospu_trace(spu_yield__enter, ctx);
 900	if (!(ctx->flags & SPU_CREATE_NOSCHED)) {
 901		mutex_lock(&ctx->state_mutex);
 902		__spu_deactivate(ctx, 0, MAX_PRIO);
 903		mutex_unlock(&ctx->state_mutex);
 904	}
 905}
 906
 907static noinline void spusched_tick(struct spu_context *ctx)
 908{
 909	struct spu_context *new = NULL;
 910	struct spu *spu = NULL;
 911
 912	if (spu_acquire(ctx))
 913		BUG();	/* a kernel thread never has signals pending */
 914
 915	if (ctx->state != SPU_STATE_RUNNABLE)
 916		goto out;
 917	if (ctx->flags & SPU_CREATE_NOSCHED)
 918		goto out;
 919	if (ctx->policy == SCHED_FIFO)
 920		goto out;
 921
 922	if (--ctx->time_slice && test_bit(SPU_SCHED_SPU_RUN, &ctx->sched_flags))
 923		goto out;
 924
 925	spu = ctx->spu;
 926
 927	spu_context_trace(spusched_tick__preempt, ctx, spu);
 928
 929	new = grab_runnable_context(ctx->prio + 1, spu->node);
 930	if (new) {
 931		spu_unschedule(spu, ctx, 0);
 932		if (test_bit(SPU_SCHED_SPU_RUN, &ctx->sched_flags))
 933			spu_add_to_rq(ctx);
 934	} else {
 935		spu_context_nospu_trace(spusched_tick__newslice, ctx);
 936		if (!ctx->time_slice)
 937			ctx->time_slice++;
 938	}
 939out:
 940	spu_release(ctx);
 941
 942	if (new)
 943		spu_schedule(spu, new);
 944}
 945
 946/**
 947 * count_active_contexts - count nr of active tasks
 948 *
 949 * Return the number of tasks currently running or waiting to run.
 950 *
 951 * Note that we don't take runq_lock / list_mutex here.  Reading
 952 * a single 32bit value is atomic on powerpc, and we don't care
 953 * about memory ordering issues here.
 954 */
 955static unsigned long count_active_contexts(void)
 956{
 957	int nr_active = 0, node;
 958
 959	for (node = 0; node < MAX_NUMNODES; node++)
 960		nr_active += cbe_spu_info[node].nr_active;
 961	nr_active += spu_prio->nr_waiting;
 962
 963	return nr_active;
 964}
 965
 966/**
 967 * spu_calc_load - update the avenrun load estimates.
 968 *
 969 * No locking against reading these values from userspace, as for
 970 * the CPU loadavg code.
 971 */
 972static void spu_calc_load(void)
 973{
 974	unsigned long active_tasks; /* fixed-point */
 975
 976	active_tasks = count_active_contexts() * FIXED_1;
 977	spu_avenrun[0] = calc_load(spu_avenrun[0], EXP_1, active_tasks);
 978	spu_avenrun[1] = calc_load(spu_avenrun[1], EXP_5, active_tasks);
 979	spu_avenrun[2] = calc_load(spu_avenrun[2], EXP_15, active_tasks);
 980}
 981
 982static void spusched_wake(struct timer_list *unused)
 983{
 984	mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
 985	wake_up_process(spusched_task);
 986}
 987
 988static void spuloadavg_wake(struct timer_list *unused)
 989{
 990	mod_timer(&spuloadavg_timer, jiffies + LOAD_FREQ);
 991	spu_calc_load();
 992}
 993
 994static int spusched_thread(void *unused)
 995{
 996	struct spu *spu;
 997	int node;
 998
 999	while (!kthread_should_stop()) {
1000		set_current_state(TASK_INTERRUPTIBLE);
1001		schedule();
1002		for (node = 0; node < MAX_NUMNODES; node++) {
1003			struct mutex *mtx = &cbe_spu_info[node].list_mutex;
1004
1005			mutex_lock(mtx);
1006			list_for_each_entry(spu, &cbe_spu_info[node].spus,
1007					cbe_list) {
1008				struct spu_context *ctx = spu->ctx;
1009
1010				if (ctx) {
1011					get_spu_context(ctx);
1012					mutex_unlock(mtx);
1013					spusched_tick(ctx);
1014					mutex_lock(mtx);
1015					put_spu_context(ctx);
1016				}
1017			}
1018			mutex_unlock(mtx);
1019		}
1020	}
1021
1022	return 0;
1023}
1024
1025void spuctx_switch_state(struct spu_context *ctx,
1026		enum spu_utilization_state new_state)
1027{
1028	unsigned long long curtime;
1029	signed long long delta;
1030	struct spu *spu;
1031	enum spu_utilization_state old_state;
1032	int node;
1033
1034	curtime = ktime_get_ns();
1035	delta = curtime - ctx->stats.tstamp;
1036
1037	WARN_ON(!mutex_is_locked(&ctx->state_mutex));
1038	WARN_ON(delta < 0);
1039
1040	spu = ctx->spu;
1041	old_state = ctx->stats.util_state;
1042	ctx->stats.util_state = new_state;
1043	ctx->stats.tstamp = curtime;
1044
1045	/*
1046	 * Update the physical SPU utilization statistics.
1047	 */
1048	if (spu) {
1049		ctx->stats.times[old_state] += delta;
1050		spu->stats.times[old_state] += delta;
1051		spu->stats.util_state = new_state;
1052		spu->stats.tstamp = curtime;
1053		node = spu->node;
1054		if (old_state == SPU_UTIL_USER)
1055			atomic_dec(&cbe_spu_info[node].busy_spus);
1056		if (new_state == SPU_UTIL_USER)
1057			atomic_inc(&cbe_spu_info[node].busy_spus);
1058	}
1059}
1060
1061static int show_spu_loadavg(struct seq_file *s, void *private)
1062{
1063	int a, b, c;
1064
1065	a = spu_avenrun[0] + (FIXED_1/200);
1066	b = spu_avenrun[1] + (FIXED_1/200);
1067	c = spu_avenrun[2] + (FIXED_1/200);
1068
1069	/*
1070	 * Note that last_pid doesn't really make much sense for the
1071	 * SPU loadavg (it even seems very odd on the CPU side...),
1072	 * but we include it here to have a 100% compatible interface.
1073	 */
1074	seq_printf(s, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n",
1075		LOAD_INT(a), LOAD_FRAC(a),
1076		LOAD_INT(b), LOAD_FRAC(b),
1077		LOAD_INT(c), LOAD_FRAC(c),
1078		count_active_contexts(),
1079		atomic_read(&nr_spu_contexts),
1080		idr_get_cursor(&task_active_pid_ns(current)->idr) - 1);
1081	return 0;
1082};
1083
1084int __init spu_sched_init(void)
1085{
1086	struct proc_dir_entry *entry;
1087	int err = -ENOMEM, i;
1088
1089	spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
1090	if (!spu_prio)
1091		goto out;
1092
1093	for (i = 0; i < MAX_PRIO; i++) {
1094		INIT_LIST_HEAD(&spu_prio->runq[i]);
1095		__clear_bit(i, spu_prio->bitmap);
1096	}
1097	spin_lock_init(&spu_prio->runq_lock);
1098
1099	timer_setup(&spusched_timer, spusched_wake, 0);
1100	timer_setup(&spuloadavg_timer, spuloadavg_wake, 0);
1101
1102	spusched_task = kthread_run(spusched_thread, NULL, "spusched");
1103	if (IS_ERR(spusched_task)) {
1104		err = PTR_ERR(spusched_task);
1105		goto out_free_spu_prio;
1106	}
1107
1108	mod_timer(&spuloadavg_timer, 0);
1109
1110	entry = proc_create_single("spu_loadavg", 0, NULL, show_spu_loadavg);
1111	if (!entry)
1112		goto out_stop_kthread;
1113
1114	pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n",
1115			SPUSCHED_TICK, MIN_SPU_TIMESLICE, DEF_SPU_TIMESLICE);
1116	return 0;
1117
1118 out_stop_kthread:
1119	kthread_stop(spusched_task);
1120 out_free_spu_prio:
1121	kfree(spu_prio);
1122 out:
1123	return err;
1124}
1125
1126void spu_sched_exit(void)
1127{
1128	struct spu *spu;
1129	int node;
1130
1131	remove_proc_entry("spu_loadavg", NULL);
1132
1133	del_timer_sync(&spusched_timer);
1134	del_timer_sync(&spuloadavg_timer);
1135	kthread_stop(spusched_task);
1136
1137	for (node = 0; node < MAX_NUMNODES; node++) {
1138		mutex_lock(&cbe_spu_info[node].list_mutex);
1139		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list)
1140			if (spu->alloc_state != SPU_FREE)
1141				spu->alloc_state = SPU_FREE;
1142		mutex_unlock(&cbe_spu_info[node].list_mutex);
1143	}
1144	kfree(spu_prio);
1145}