1 /* sched.c - SPU scheduler.
2  *
3  * Copyright (C) IBM 2005
4  * Author: Mark Nutter <mnutter@us.ibm.com>
5  *
6  * 2006-03-31	NUMA domains added.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License as published by
10  * the Free Software Foundation; either version 2, or (at your option)
11  * any later version.
12  *
13  * This program is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16  * GNU General Public License for more details.
17  *
18  * You should have received a copy of the GNU General Public License
19  * along with this program; if not, write to the Free Software
20  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
21  */
22 
23 #undef DEBUG
24 
25 #include <linux/errno.h>
26 #include <linux/sched.h>
27 #include <linux/sched/rt.h>
28 #include <linux/kernel.h>
29 #include <linux/mm.h>
30 #include <linux/slab.h>
31 #include <linux/completion.h>
32 #include <linux/vmalloc.h>
33 #include <linux/smp.h>
34 #include <linux/stddef.h>
35 #include <linux/unistd.h>
36 #include <linux/numa.h>
37 #include <linux/mutex.h>
38 #include <linux/notifier.h>
39 #include <linux/kthread.h>
40 #include <linux/pid_namespace.h>
41 #include <linux/proc_fs.h>
42 #include <linux/seq_file.h>
43 
44 #include <asm/io.h>
45 #include <asm/mmu_context.h>
46 #include <asm/spu.h>
47 #include <asm/spu_csa.h>
48 #include <asm/spu_priv1.h>
49 #include "spufs.h"
50 #define CREATE_TRACE_POINTS
51 #include "sputrace.h"
52 
53 struct spu_prio_array {
54 	DECLARE_BITMAP(bitmap, MAX_PRIO);
55 	struct list_head runq[MAX_PRIO];
56 	spinlock_t runq_lock;
57 	int nr_waiting;
58 };
59 
60 static unsigned long spu_avenrun[3];
61 static struct spu_prio_array *spu_prio;
62 static struct task_struct *spusched_task;
63 static struct timer_list spusched_timer;
64 static struct timer_list spuloadavg_timer;
65 
66 /*
67  * Priority of a normal, non-rt, non-niced'd process (aka nice level 0).
68  */
69 #define NORMAL_PRIO		120
70 
71 /*
72  * Frequency of the spu scheduler tick.  By default we do one SPU scheduler
73  * tick for every 10 CPU scheduler ticks.
74  */
75 #define SPUSCHED_TICK		(10)
76 
77 /*
78  * These are the 'tuning knobs' of the scheduler:
79  *
80  * Minimum timeslice is 5 msecs (or 1 spu scheduler tick, whichever is
81  * larger), default timeslice is 100 msecs, maximum timeslice is 800 msecs.
82  */
83 #define MIN_SPU_TIMESLICE	max(5 * HZ / (1000 * SPUSCHED_TICK), 1)
84 #define DEF_SPU_TIMESLICE	(100 * HZ / (1000 * SPUSCHED_TICK))
85 
86 #define SCALE_PRIO(x, prio) \
87 	max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_SPU_TIMESLICE)
88 
89 /*
90  * scale user-nice values [ -20 ... 0 ... 19 ] to time slice values:
91  * [800ms ... 100ms ... 5ms]
92  *
93  * The higher a thread's priority, the bigger timeslices
94  * it gets during one round of execution. But even the lowest
95  * priority thread gets MIN_TIMESLICE worth of execution time.
96  */
97 void spu_set_timeslice(struct spu_context *ctx)
98 {
99 	if (ctx->prio < NORMAL_PRIO)
100 		ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE * 4, ctx->prio);
101 	else
102 		ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE, ctx->prio);
103 }
104 
105 /*
106  * Update scheduling information from the owning thread.
107  */
108 void __spu_update_sched_info(struct spu_context *ctx)
109 {
110 	/*
111 	 * assert that the context is not on the runqueue, so it is safe
112 	 * to change its scheduling parameters.
113 	 */
114 	BUG_ON(!list_empty(&ctx->rq));
115 
116 	/*
117 	 * 32-Bit assignments are atomic on powerpc, and we don't care about
118 	 * memory ordering here because retrieving the controlling thread is
119 	 * per definition racy.
120 	 */
121 	ctx->tid = current->pid;
122 
123 	/*
124 	 * We do our own priority calculations, so we normally want
125 	 * ->static_prio to start with. Unfortunately this field
126 	 * contains junk for threads with a realtime scheduling
127 	 * policy so we have to look at ->prio in this case.
128 	 */
129 	if (rt_prio(current->prio))
130 		ctx->prio = current->prio;
131 	else
132 		ctx->prio = current->static_prio;
133 	ctx->policy = current->policy;
134 
135 	/*
136 	 * TO DO: the context may be loaded, so we may need to activate
137 	 * it again on a different node. But it shouldn't hurt anything
138 	 * to update its parameters, because we know that the scheduler
139 	 * is not actively looking at this field, since it is not on the
140 	 * runqueue. The context will be rescheduled on the proper node
141 	 * if it is timesliced or preempted.
142 	 */
143 	cpumask_copy(&ctx->cpus_allowed, tsk_cpus_allowed(current));
144 
145 	/* Save the current cpu id for spu interrupt routing. */
146 	ctx->last_ran = raw_smp_processor_id();
147 }
148 
149 void spu_update_sched_info(struct spu_context *ctx)
150 {
151 	int node;
152 
153 	if (ctx->state == SPU_STATE_RUNNABLE) {
154 		node = ctx->spu->node;
155 
156 		/*
157 		 * Take list_mutex to sync with find_victim().
158 		 */
159 		mutex_lock(&cbe_spu_info[node].list_mutex);
160 		__spu_update_sched_info(ctx);
161 		mutex_unlock(&cbe_spu_info[node].list_mutex);
162 	} else {
163 		__spu_update_sched_info(ctx);
164 	}
165 }
166 
167 static int __node_allowed(struct spu_context *ctx, int node)
168 {
169 	if (nr_cpus_node(node)) {
170 		const struct cpumask *mask = cpumask_of_node(node);
171 
172 		if (cpumask_intersects(mask, &ctx->cpus_allowed))
173 			return 1;
174 	}
175 
176 	return 0;
177 }
178 
179 static int node_allowed(struct spu_context *ctx, int node)
180 {
181 	int rval;
182 
183 	spin_lock(&spu_prio->runq_lock);
184 	rval = __node_allowed(ctx, node);
185 	spin_unlock(&spu_prio->runq_lock);
186 
187 	return rval;
188 }
189 
190 void do_notify_spus_active(void)
191 {
192 	int node;
193 
194 	/*
195 	 * Wake up the active spu_contexts.
196 	 *
197 	 * When the awakened processes see their "notify_active" flag is set,
198 	 * they will call spu_switch_notify().
199 	 */
200 	for_each_online_node(node) {
201 		struct spu *spu;
202 
203 		mutex_lock(&cbe_spu_info[node].list_mutex);
204 		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
205 			if (spu->alloc_state != SPU_FREE) {
206 				struct spu_context *ctx = spu->ctx;
207 				set_bit(SPU_SCHED_NOTIFY_ACTIVE,
208 					&ctx->sched_flags);
209 				mb();
210 				wake_up_all(&ctx->stop_wq);
211 			}
212 		}
213 		mutex_unlock(&cbe_spu_info[node].list_mutex);
214 	}
215 }
216 
217 /**
218  * spu_bind_context - bind spu context to physical spu
219  * @spu:	physical spu to bind to
220  * @ctx:	context to bind
221  */
222 static void spu_bind_context(struct spu *spu, struct spu_context *ctx)
223 {
224 	spu_context_trace(spu_bind_context__enter, ctx, spu);
225 
226 	spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
227 
228 	if (ctx->flags & SPU_CREATE_NOSCHED)
229 		atomic_inc(&cbe_spu_info[spu->node].reserved_spus);
230 
231 	ctx->stats.slb_flt_base = spu->stats.slb_flt;
232 	ctx->stats.class2_intr_base = spu->stats.class2_intr;
233 
234 	spu_associate_mm(spu, ctx->owner);
235 
236 	spin_lock_irq(&spu->register_lock);
237 	spu->ctx = ctx;
238 	spu->flags = 0;
239 	ctx->spu = spu;
240 	ctx->ops = &spu_hw_ops;
241 	spu->pid = current->pid;
242 	spu->tgid = current->tgid;
243 	spu->ibox_callback = spufs_ibox_callback;
244 	spu->wbox_callback = spufs_wbox_callback;
245 	spu->stop_callback = spufs_stop_callback;
246 	spu->mfc_callback = spufs_mfc_callback;
247 	spin_unlock_irq(&spu->register_lock);
248 
249 	spu_unmap_mappings(ctx);
250 
251 	spu_switch_log_notify(spu, ctx, SWITCH_LOG_START, 0);
252 	spu_restore(&ctx->csa, spu);
253 	spu->timestamp = jiffies;
254 	spu_switch_notify(spu, ctx);
255 	ctx->state = SPU_STATE_RUNNABLE;
256 
257 	spuctx_switch_state(ctx, SPU_UTIL_USER);
258 }
259 
260 /*
261  * Must be used with the list_mutex held.
262  */
263 static inline int sched_spu(struct spu *spu)
264 {
265 	BUG_ON(!mutex_is_locked(&cbe_spu_info[spu->node].list_mutex));
266 
267 	return (!spu->ctx || !(spu->ctx->flags & SPU_CREATE_NOSCHED));
268 }
269 
270 static void aff_merge_remaining_ctxs(struct spu_gang *gang)
271 {
272 	struct spu_context *ctx;
273 
274 	list_for_each_entry(ctx, &gang->aff_list_head, aff_list) {
275 		if (list_empty(&ctx->aff_list))
276 			list_add(&ctx->aff_list, &gang->aff_list_head);
277 	}
278 	gang->aff_flags |= AFF_MERGED;
279 }
280 
281 static void aff_set_offsets(struct spu_gang *gang)
282 {
283 	struct spu_context *ctx;
284 	int offset;
285 
286 	offset = -1;
287 	list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
288 								aff_list) {
289 		if (&ctx->aff_list == &gang->aff_list_head)
290 			break;
291 		ctx->aff_offset = offset--;
292 	}
293 
294 	offset = 0;
295 	list_for_each_entry(ctx, gang->aff_ref_ctx->aff_list.prev, aff_list) {
296 		if (&ctx->aff_list == &gang->aff_list_head)
297 			break;
298 		ctx->aff_offset = offset++;
299 	}
300 
301 	gang->aff_flags |= AFF_OFFSETS_SET;
302 }
303 
304 static struct spu *aff_ref_location(struct spu_context *ctx, int mem_aff,
305 		 int group_size, int lowest_offset)
306 {
307 	struct spu *spu;
308 	int node, n;
309 
310 	/*
311 	 * TODO: A better algorithm could be used to find a good spu to be
312 	 *       used as reference location for the ctxs chain.
313 	 */
314 	node = cpu_to_node(raw_smp_processor_id());
315 	for (n = 0; n < MAX_NUMNODES; n++, node++) {
316 		/*
317 		 * "available_spus" counts how many spus are not potentially
318 		 * going to be used by other affinity gangs whose reference
319 		 * context is already in place. Although this code seeks to
320 		 * avoid having affinity gangs with a summed amount of
321 		 * contexts bigger than the amount of spus in the node,
322 		 * this may happen sporadically. In this case, available_spus
323 		 * becomes negative, which is harmless.
324 		 */
325 		int available_spus;
326 
327 		node = (node < MAX_NUMNODES) ? node : 0;
328 		if (!node_allowed(ctx, node))
329 			continue;
330 
331 		available_spus = 0;
332 		mutex_lock(&cbe_spu_info[node].list_mutex);
333 		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
334 			if (spu->ctx && spu->ctx->gang && !spu->ctx->aff_offset
335 					&& spu->ctx->gang->aff_ref_spu)
336 				available_spus -= spu->ctx->gang->contexts;
337 			available_spus++;
338 		}
339 		if (available_spus < ctx->gang->contexts) {
340 			mutex_unlock(&cbe_spu_info[node].list_mutex);
341 			continue;
342 		}
343 
344 		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
345 			if ((!mem_aff || spu->has_mem_affinity) &&
346 							sched_spu(spu)) {
347 				mutex_unlock(&cbe_spu_info[node].list_mutex);
348 				return spu;
349 			}
350 		}
351 		mutex_unlock(&cbe_spu_info[node].list_mutex);
352 	}
353 	return NULL;
354 }
355 
356 static void aff_set_ref_point_location(struct spu_gang *gang)
357 {
358 	int mem_aff, gs, lowest_offset;
359 	struct spu_context *ctx;
360 	struct spu *tmp;
361 
362 	mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM;
363 	lowest_offset = 0;
364 	gs = 0;
365 
366 	list_for_each_entry(tmp, &gang->aff_list_head, aff_list)
367 		gs++;
368 
369 	list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
370 								aff_list) {
371 		if (&ctx->aff_list == &gang->aff_list_head)
372 			break;
373 		lowest_offset = ctx->aff_offset;
374 	}
375 
376 	gang->aff_ref_spu = aff_ref_location(gang->aff_ref_ctx, mem_aff, gs,
377 							lowest_offset);
378 }
379 
380 static struct spu *ctx_location(struct spu *ref, int offset, int node)
381 {
382 	struct spu *spu;
383 
384 	spu = NULL;
385 	if (offset >= 0) {
386 		list_for_each_entry(spu, ref->aff_list.prev, aff_list) {
387 			BUG_ON(spu->node != node);
388 			if (offset == 0)
389 				break;
390 			if (sched_spu(spu))
391 				offset--;
392 		}
393 	} else {
394 		list_for_each_entry_reverse(spu, ref->aff_list.next, aff_list) {
395 			BUG_ON(spu->node != node);
396 			if (offset == 0)
397 				break;
398 			if (sched_spu(spu))
399 				offset++;
400 		}
401 	}
402 
403 	return spu;
404 }
405 
406 /*
407  * affinity_check is called each time a context is going to be scheduled.
408  * It returns the spu ptr on which the context must run.
409  */
410 static int has_affinity(struct spu_context *ctx)
411 {
412 	struct spu_gang *gang = ctx->gang;
413 
414 	if (list_empty(&ctx->aff_list))
415 		return 0;
416 
417 	if (atomic_read(&ctx->gang->aff_sched_count) == 0)
418 		ctx->gang->aff_ref_spu = NULL;
419 
420 	if (!gang->aff_ref_spu) {
421 		if (!(gang->aff_flags & AFF_MERGED))
422 			aff_merge_remaining_ctxs(gang);
423 		if (!(gang->aff_flags & AFF_OFFSETS_SET))
424 			aff_set_offsets(gang);
425 		aff_set_ref_point_location(gang);
426 	}
427 
428 	return gang->aff_ref_spu != NULL;
429 }
430 
431 /**
432  * spu_unbind_context - unbind spu context from physical spu
433  * @spu:	physical spu to unbind from
434  * @ctx:	context to unbind
435  */
436 static void spu_unbind_context(struct spu *spu, struct spu_context *ctx)
437 {
438 	u32 status;
439 
440 	spu_context_trace(spu_unbind_context__enter, ctx, spu);
441 
442 	spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
443 
444  	if (spu->ctx->flags & SPU_CREATE_NOSCHED)
445 		atomic_dec(&cbe_spu_info[spu->node].reserved_spus);
446 
447 	if (ctx->gang)
448 		/*
449 		 * If ctx->gang->aff_sched_count is positive, SPU affinity is
450 		 * being considered in this gang. Using atomic_dec_if_positive
451 		 * allow us to skip an explicit check for affinity in this gang
452 		 */
453 		atomic_dec_if_positive(&ctx->gang->aff_sched_count);
454 
455 	spu_switch_notify(spu, NULL);
456 	spu_unmap_mappings(ctx);
457 	spu_save(&ctx->csa, spu);
458 	spu_switch_log_notify(spu, ctx, SWITCH_LOG_STOP, 0);
459 
460 	spin_lock_irq(&spu->register_lock);
461 	spu->timestamp = jiffies;
462 	ctx->state = SPU_STATE_SAVED;
463 	spu->ibox_callback = NULL;
464 	spu->wbox_callback = NULL;
465 	spu->stop_callback = NULL;
466 	spu->mfc_callback = NULL;
467 	spu->pid = 0;
468 	spu->tgid = 0;
469 	ctx->ops = &spu_backing_ops;
470 	spu->flags = 0;
471 	spu->ctx = NULL;
472 	spin_unlock_irq(&spu->register_lock);
473 
474 	spu_associate_mm(spu, NULL);
475 
476 	ctx->stats.slb_flt +=
477 		(spu->stats.slb_flt - ctx->stats.slb_flt_base);
478 	ctx->stats.class2_intr +=
479 		(spu->stats.class2_intr - ctx->stats.class2_intr_base);
480 
481 	/* This maps the underlying spu state to idle */
482 	spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED);
483 	ctx->spu = NULL;
484 
485 	if (spu_stopped(ctx, &status))
486 		wake_up_all(&ctx->stop_wq);
487 }
488 
489 /**
490  * spu_add_to_rq - add a context to the runqueue
491  * @ctx:       context to add
492  */
493 static void __spu_add_to_rq(struct spu_context *ctx)
494 {
495 	/*
496 	 * Unfortunately this code path can be called from multiple threads
497 	 * on behalf of a single context due to the way the problem state
498 	 * mmap support works.
499 	 *
500 	 * Fortunately we need to wake up all these threads at the same time
501 	 * and can simply skip the runqueue addition for every but the first
502 	 * thread getting into this codepath.
503 	 *
504 	 * It's still quite hacky, and long-term we should proxy all other
505 	 * threads through the owner thread so that spu_run is in control
506 	 * of all the scheduling activity for a given context.
507 	 */
508 	if (list_empty(&ctx->rq)) {
509 		list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]);
510 		set_bit(ctx->prio, spu_prio->bitmap);
511 		if (!spu_prio->nr_waiting++)
512 			mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
513 	}
514 }
515 
516 static void spu_add_to_rq(struct spu_context *ctx)
517 {
518 	spin_lock(&spu_prio->runq_lock);
519 	__spu_add_to_rq(ctx);
520 	spin_unlock(&spu_prio->runq_lock);
521 }
522 
523 static void __spu_del_from_rq(struct spu_context *ctx)
524 {
525 	int prio = ctx->prio;
526 
527 	if (!list_empty(&ctx->rq)) {
528 		if (!--spu_prio->nr_waiting)
529 			del_timer(&spusched_timer);
530 		list_del_init(&ctx->rq);
531 
532 		if (list_empty(&spu_prio->runq[prio]))
533 			clear_bit(prio, spu_prio->bitmap);
534 	}
535 }
536 
537 void spu_del_from_rq(struct spu_context *ctx)
538 {
539 	spin_lock(&spu_prio->runq_lock);
540 	__spu_del_from_rq(ctx);
541 	spin_unlock(&spu_prio->runq_lock);
542 }
543 
544 static void spu_prio_wait(struct spu_context *ctx)
545 {
546 	DEFINE_WAIT(wait);
547 
548 	/*
549 	 * The caller must explicitly wait for a context to be loaded
550 	 * if the nosched flag is set.  If NOSCHED is not set, the caller
551 	 * queues the context and waits for an spu event or error.
552 	 */
553 	BUG_ON(!(ctx->flags & SPU_CREATE_NOSCHED));
554 
555 	spin_lock(&spu_prio->runq_lock);
556 	prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE);
557 	if (!signal_pending(current)) {
558 		__spu_add_to_rq(ctx);
559 		spin_unlock(&spu_prio->runq_lock);
560 		mutex_unlock(&ctx->state_mutex);
561 		schedule();
562 		mutex_lock(&ctx->state_mutex);
563 		spin_lock(&spu_prio->runq_lock);
564 		__spu_del_from_rq(ctx);
565 	}
566 	spin_unlock(&spu_prio->runq_lock);
567 	__set_current_state(TASK_RUNNING);
568 	remove_wait_queue(&ctx->stop_wq, &wait);
569 }
570 
571 static struct spu *spu_get_idle(struct spu_context *ctx)
572 {
573 	struct spu *spu, *aff_ref_spu;
574 	int node, n;
575 
576 	spu_context_nospu_trace(spu_get_idle__enter, ctx);
577 
578 	if (ctx->gang) {
579 		mutex_lock(&ctx->gang->aff_mutex);
580 		if (has_affinity(ctx)) {
581 			aff_ref_spu = ctx->gang->aff_ref_spu;
582 			atomic_inc(&ctx->gang->aff_sched_count);
583 			mutex_unlock(&ctx->gang->aff_mutex);
584 			node = aff_ref_spu->node;
585 
586 			mutex_lock(&cbe_spu_info[node].list_mutex);
587 			spu = ctx_location(aff_ref_spu, ctx->aff_offset, node);
588 			if (spu && spu->alloc_state == SPU_FREE)
589 				goto found;
590 			mutex_unlock(&cbe_spu_info[node].list_mutex);
591 
592 			atomic_dec(&ctx->gang->aff_sched_count);
593 			goto not_found;
594 		}
595 		mutex_unlock(&ctx->gang->aff_mutex);
596 	}
597 	node = cpu_to_node(raw_smp_processor_id());
598 	for (n = 0; n < MAX_NUMNODES; n++, node++) {
599 		node = (node < MAX_NUMNODES) ? node : 0;
600 		if (!node_allowed(ctx, node))
601 			continue;
602 
603 		mutex_lock(&cbe_spu_info[node].list_mutex);
604 		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
605 			if (spu->alloc_state == SPU_FREE)
606 				goto found;
607 		}
608 		mutex_unlock(&cbe_spu_info[node].list_mutex);
609 	}
610 
611  not_found:
612 	spu_context_nospu_trace(spu_get_idle__not_found, ctx);
613 	return NULL;
614 
615  found:
616 	spu->alloc_state = SPU_USED;
617 	mutex_unlock(&cbe_spu_info[node].list_mutex);
618 	spu_context_trace(spu_get_idle__found, ctx, spu);
619 	spu_init_channels(spu);
620 	return spu;
621 }
622 
623 /**
624  * find_victim - find a lower priority context to preempt
625  * @ctx:	candidate context for running
626  *
627  * Returns the freed physical spu to run the new context on.
628  */
629 static struct spu *find_victim(struct spu_context *ctx)
630 {
631 	struct spu_context *victim = NULL;
632 	struct spu *spu;
633 	int node, n;
634 
635 	spu_context_nospu_trace(spu_find_victim__enter, ctx);
636 
637 	/*
638 	 * Look for a possible preemption candidate on the local node first.
639 	 * If there is no candidate look at the other nodes.  This isn't
640 	 * exactly fair, but so far the whole spu scheduler tries to keep
641 	 * a strong node affinity.  We might want to fine-tune this in
642 	 * the future.
643 	 */
644  restart:
645 	node = cpu_to_node(raw_smp_processor_id());
646 	for (n = 0; n < MAX_NUMNODES; n++, node++) {
647 		node = (node < MAX_NUMNODES) ? node : 0;
648 		if (!node_allowed(ctx, node))
649 			continue;
650 
651 		mutex_lock(&cbe_spu_info[node].list_mutex);
652 		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
653 			struct spu_context *tmp = spu->ctx;
654 
655 			if (tmp && tmp->prio > ctx->prio &&
656 			    !(tmp->flags & SPU_CREATE_NOSCHED) &&
657 			    (!victim || tmp->prio > victim->prio)) {
658 				victim = spu->ctx;
659 			}
660 		}
661 		if (victim)
662 			get_spu_context(victim);
663 		mutex_unlock(&cbe_spu_info[node].list_mutex);
664 
665 		if (victim) {
666 			/*
667 			 * This nests ctx->state_mutex, but we always lock
668 			 * higher priority contexts before lower priority
669 			 * ones, so this is safe until we introduce
670 			 * priority inheritance schemes.
671 			 *
672 			 * XXX if the highest priority context is locked,
673 			 * this can loop a long time.  Might be better to
674 			 * look at another context or give up after X retries.
675 			 */
676 			if (!mutex_trylock(&victim->state_mutex)) {
677 				put_spu_context(victim);
678 				victim = NULL;
679 				goto restart;
680 			}
681 
682 			spu = victim->spu;
683 			if (!spu || victim->prio <= ctx->prio) {
684 				/*
685 				 * This race can happen because we've dropped
686 				 * the active list mutex.  Not a problem, just
687 				 * restart the search.
688 				 */
689 				mutex_unlock(&victim->state_mutex);
690 				put_spu_context(victim);
691 				victim = NULL;
692 				goto restart;
693 			}
694 
695 			spu_context_trace(__spu_deactivate__unload, ctx, spu);
696 
697 			mutex_lock(&cbe_spu_info[node].list_mutex);
698 			cbe_spu_info[node].nr_active--;
699 			spu_unbind_context(spu, victim);
700 			mutex_unlock(&cbe_spu_info[node].list_mutex);
701 
702 			victim->stats.invol_ctx_switch++;
703 			spu->stats.invol_ctx_switch++;
704 			if (test_bit(SPU_SCHED_SPU_RUN, &victim->sched_flags))
705 				spu_add_to_rq(victim);
706 
707 			mutex_unlock(&victim->state_mutex);
708 			put_spu_context(victim);
709 
710 			return spu;
711 		}
712 	}
713 
714 	return NULL;
715 }
716 
717 static void __spu_schedule(struct spu *spu, struct spu_context *ctx)
718 {
719 	int node = spu->node;
720 	int success = 0;
721 
722 	spu_set_timeslice(ctx);
723 
724 	mutex_lock(&cbe_spu_info[node].list_mutex);
725 	if (spu->ctx == NULL) {
726 		spu_bind_context(spu, ctx);
727 		cbe_spu_info[node].nr_active++;
728 		spu->alloc_state = SPU_USED;
729 		success = 1;
730 	}
731 	mutex_unlock(&cbe_spu_info[node].list_mutex);
732 
733 	if (success)
734 		wake_up_all(&ctx->run_wq);
735 	else
736 		spu_add_to_rq(ctx);
737 }
738 
739 static void spu_schedule(struct spu *spu, struct spu_context *ctx)
740 {
741 	/* not a candidate for interruptible because it's called either
742 	   from the scheduler thread or from spu_deactivate */
743 	mutex_lock(&ctx->state_mutex);
744 	if (ctx->state == SPU_STATE_SAVED)
745 		__spu_schedule(spu, ctx);
746 	spu_release(ctx);
747 }
748 
749 /**
750  * spu_unschedule - remove a context from a spu, and possibly release it.
751  * @spu:	The SPU to unschedule from
752  * @ctx:	The context currently scheduled on the SPU
753  * @free_spu	Whether to free the SPU for other contexts
754  *
755  * Unbinds the context @ctx from the SPU @spu. If @free_spu is non-zero, the
756  * SPU is made available for other contexts (ie, may be returned by
757  * spu_get_idle). If this is zero, the caller is expected to schedule another
758  * context to this spu.
759  *
760  * Should be called with ctx->state_mutex held.
761  */
762 static void spu_unschedule(struct spu *spu, struct spu_context *ctx,
763 		int free_spu)
764 {
765 	int node = spu->node;
766 
767 	mutex_lock(&cbe_spu_info[node].list_mutex);
768 	cbe_spu_info[node].nr_active--;
769 	if (free_spu)
770 		spu->alloc_state = SPU_FREE;
771 	spu_unbind_context(spu, ctx);
772 	ctx->stats.invol_ctx_switch++;
773 	spu->stats.invol_ctx_switch++;
774 	mutex_unlock(&cbe_spu_info[node].list_mutex);
775 }
776 
777 /**
778  * spu_activate - find a free spu for a context and execute it
779  * @ctx:	spu context to schedule
780  * @flags:	flags (currently ignored)
781  *
782  * Tries to find a free spu to run @ctx.  If no free spu is available
783  * add the context to the runqueue so it gets woken up once an spu
784  * is available.
785  */
786 int spu_activate(struct spu_context *ctx, unsigned long flags)
787 {
788 	struct spu *spu;
789 
790 	/*
791 	 * If there are multiple threads waiting for a single context
792 	 * only one actually binds the context while the others will
793 	 * only be able to acquire the state_mutex once the context
794 	 * already is in runnable state.
795 	 */
796 	if (ctx->spu)
797 		return 0;
798 
799 spu_activate_top:
800 	if (signal_pending(current))
801 		return -ERESTARTSYS;
802 
803 	spu = spu_get_idle(ctx);
804 	/*
805 	 * If this is a realtime thread we try to get it running by
806 	 * preempting a lower priority thread.
807 	 */
808 	if (!spu && rt_prio(ctx->prio))
809 		spu = find_victim(ctx);
810 	if (spu) {
811 		unsigned long runcntl;
812 
813 		runcntl = ctx->ops->runcntl_read(ctx);
814 		__spu_schedule(spu, ctx);
815 		if (runcntl & SPU_RUNCNTL_RUNNABLE)
816 			spuctx_switch_state(ctx, SPU_UTIL_USER);
817 
818 		return 0;
819 	}
820 
821 	if (ctx->flags & SPU_CREATE_NOSCHED) {
822 		spu_prio_wait(ctx);
823 		goto spu_activate_top;
824 	}
825 
826 	spu_add_to_rq(ctx);
827 
828 	return 0;
829 }
830 
831 /**
832  * grab_runnable_context - try to find a runnable context
833  *
834  * Remove the highest priority context on the runqueue and return it
835  * to the caller.  Returns %NULL if no runnable context was found.
836  */
837 static struct spu_context *grab_runnable_context(int prio, int node)
838 {
839 	struct spu_context *ctx;
840 	int best;
841 
842 	spin_lock(&spu_prio->runq_lock);
843 	best = find_first_bit(spu_prio->bitmap, prio);
844 	while (best < prio) {
845 		struct list_head *rq = &spu_prio->runq[best];
846 
847 		list_for_each_entry(ctx, rq, rq) {
848 			/* XXX(hch): check for affinity here as well */
849 			if (__node_allowed(ctx, node)) {
850 				__spu_del_from_rq(ctx);
851 				goto found;
852 			}
853 		}
854 		best++;
855 	}
856 	ctx = NULL;
857  found:
858 	spin_unlock(&spu_prio->runq_lock);
859 	return ctx;
860 }
861 
862 static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio)
863 {
864 	struct spu *spu = ctx->spu;
865 	struct spu_context *new = NULL;
866 
867 	if (spu) {
868 		new = grab_runnable_context(max_prio, spu->node);
869 		if (new || force) {
870 			spu_unschedule(spu, ctx, new == NULL);
871 			if (new) {
872 				if (new->flags & SPU_CREATE_NOSCHED)
873 					wake_up(&new->stop_wq);
874 				else {
875 					spu_release(ctx);
876 					spu_schedule(spu, new);
877 					/* this one can't easily be made
878 					   interruptible */
879 					mutex_lock(&ctx->state_mutex);
880 				}
881 			}
882 		}
883 	}
884 
885 	return new != NULL;
886 }
887 
888 /**
889  * spu_deactivate - unbind a context from it's physical spu
890  * @ctx:	spu context to unbind
891  *
892  * Unbind @ctx from the physical spu it is running on and schedule
893  * the highest priority context to run on the freed physical spu.
894  */
895 void spu_deactivate(struct spu_context *ctx)
896 {
897 	spu_context_nospu_trace(spu_deactivate__enter, ctx);
898 	__spu_deactivate(ctx, 1, MAX_PRIO);
899 }
900 
901 /**
902  * spu_yield -	yield a physical spu if others are waiting
903  * @ctx:	spu context to yield
904  *
905  * Check if there is a higher priority context waiting and if yes
906  * unbind @ctx from the physical spu and schedule the highest
907  * priority context to run on the freed physical spu instead.
908  */
909 void spu_yield(struct spu_context *ctx)
910 {
911 	spu_context_nospu_trace(spu_yield__enter, ctx);
912 	if (!(ctx->flags & SPU_CREATE_NOSCHED)) {
913 		mutex_lock(&ctx->state_mutex);
914 		__spu_deactivate(ctx, 0, MAX_PRIO);
915 		mutex_unlock(&ctx->state_mutex);
916 	}
917 }
918 
919 static noinline void spusched_tick(struct spu_context *ctx)
920 {
921 	struct spu_context *new = NULL;
922 	struct spu *spu = NULL;
923 
924 	if (spu_acquire(ctx))
925 		BUG();	/* a kernel thread never has signals pending */
926 
927 	if (ctx->state != SPU_STATE_RUNNABLE)
928 		goto out;
929 	if (ctx->flags & SPU_CREATE_NOSCHED)
930 		goto out;
931 	if (ctx->policy == SCHED_FIFO)
932 		goto out;
933 
934 	if (--ctx->time_slice && test_bit(SPU_SCHED_SPU_RUN, &ctx->sched_flags))
935 		goto out;
936 
937 	spu = ctx->spu;
938 
939 	spu_context_trace(spusched_tick__preempt, ctx, spu);
940 
941 	new = grab_runnable_context(ctx->prio + 1, spu->node);
942 	if (new) {
943 		spu_unschedule(spu, ctx, 0);
944 		if (test_bit(SPU_SCHED_SPU_RUN, &ctx->sched_flags))
945 			spu_add_to_rq(ctx);
946 	} else {
947 		spu_context_nospu_trace(spusched_tick__newslice, ctx);
948 		if (!ctx->time_slice)
949 			ctx->time_slice++;
950 	}
951 out:
952 	spu_release(ctx);
953 
954 	if (new)
955 		spu_schedule(spu, new);
956 }
957 
958 /**
959  * count_active_contexts - count nr of active tasks
960  *
961  * Return the number of tasks currently running or waiting to run.
962  *
963  * Note that we don't take runq_lock / list_mutex here.  Reading
964  * a single 32bit value is atomic on powerpc, and we don't care
965  * about memory ordering issues here.
966  */
967 static unsigned long count_active_contexts(void)
968 {
969 	int nr_active = 0, node;
970 
971 	for (node = 0; node < MAX_NUMNODES; node++)
972 		nr_active += cbe_spu_info[node].nr_active;
973 	nr_active += spu_prio->nr_waiting;
974 
975 	return nr_active;
976 }
977 
978 /**
979  * spu_calc_load - update the avenrun load estimates.
980  *
981  * No locking against reading these values from userspace, as for
982  * the CPU loadavg code.
983  */
984 static void spu_calc_load(void)
985 {
986 	unsigned long active_tasks; /* fixed-point */
987 
988 	active_tasks = count_active_contexts() * FIXED_1;
989 	CALC_LOAD(spu_avenrun[0], EXP_1, active_tasks);
990 	CALC_LOAD(spu_avenrun[1], EXP_5, active_tasks);
991 	CALC_LOAD(spu_avenrun[2], EXP_15, active_tasks);
992 }
993 
994 static void spusched_wake(unsigned long data)
995 {
996 	mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
997 	wake_up_process(spusched_task);
998 }
999 
1000 static void spuloadavg_wake(unsigned long data)
1001 {
1002 	mod_timer(&spuloadavg_timer, jiffies + LOAD_FREQ);
1003 	spu_calc_load();
1004 }
1005 
1006 static int spusched_thread(void *unused)
1007 {
1008 	struct spu *spu;
1009 	int node;
1010 
1011 	while (!kthread_should_stop()) {
1012 		set_current_state(TASK_INTERRUPTIBLE);
1013 		schedule();
1014 		for (node = 0; node < MAX_NUMNODES; node++) {
1015 			struct mutex *mtx = &cbe_spu_info[node].list_mutex;
1016 
1017 			mutex_lock(mtx);
1018 			list_for_each_entry(spu, &cbe_spu_info[node].spus,
1019 					cbe_list) {
1020 				struct spu_context *ctx = spu->ctx;
1021 
1022 				if (ctx) {
1023 					get_spu_context(ctx);
1024 					mutex_unlock(mtx);
1025 					spusched_tick(ctx);
1026 					mutex_lock(mtx);
1027 					put_spu_context(ctx);
1028 				}
1029 			}
1030 			mutex_unlock(mtx);
1031 		}
1032 	}
1033 
1034 	return 0;
1035 }
1036 
1037 void spuctx_switch_state(struct spu_context *ctx,
1038 		enum spu_utilization_state new_state)
1039 {
1040 	unsigned long long curtime;
1041 	signed long long delta;
1042 	struct spu *spu;
1043 	enum spu_utilization_state old_state;
1044 	int node;
1045 
1046 	curtime = ktime_get_ns();
1047 	delta = curtime - ctx->stats.tstamp;
1048 
1049 	WARN_ON(!mutex_is_locked(&ctx->state_mutex));
1050 	WARN_ON(delta < 0);
1051 
1052 	spu = ctx->spu;
1053 	old_state = ctx->stats.util_state;
1054 	ctx->stats.util_state = new_state;
1055 	ctx->stats.tstamp = curtime;
1056 
1057 	/*
1058 	 * Update the physical SPU utilization statistics.
1059 	 */
1060 	if (spu) {
1061 		ctx->stats.times[old_state] += delta;
1062 		spu->stats.times[old_state] += delta;
1063 		spu->stats.util_state = new_state;
1064 		spu->stats.tstamp = curtime;
1065 		node = spu->node;
1066 		if (old_state == SPU_UTIL_USER)
1067 			atomic_dec(&cbe_spu_info[node].busy_spus);
1068 		if (new_state == SPU_UTIL_USER)
1069 			atomic_inc(&cbe_spu_info[node].busy_spus);
1070 	}
1071 }
1072 
1073 #define LOAD_INT(x) ((x) >> FSHIFT)
1074 #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
1075 
1076 static int show_spu_loadavg(struct seq_file *s, void *private)
1077 {
1078 	int a, b, c;
1079 
1080 	a = spu_avenrun[0] + (FIXED_1/200);
1081 	b = spu_avenrun[1] + (FIXED_1/200);
1082 	c = spu_avenrun[2] + (FIXED_1/200);
1083 
1084 	/*
1085 	 * Note that last_pid doesn't really make much sense for the
1086 	 * SPU loadavg (it even seems very odd on the CPU side...),
1087 	 * but we include it here to have a 100% compatible interface.
1088 	 */
1089 	seq_printf(s, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n",
1090 		LOAD_INT(a), LOAD_FRAC(a),
1091 		LOAD_INT(b), LOAD_FRAC(b),
1092 		LOAD_INT(c), LOAD_FRAC(c),
1093 		count_active_contexts(),
1094 		atomic_read(&nr_spu_contexts),
1095 		task_active_pid_ns(current)->last_pid);
1096 	return 0;
1097 }
1098 
1099 static int spu_loadavg_open(struct inode *inode, struct file *file)
1100 {
1101 	return single_open(file, show_spu_loadavg, NULL);
1102 }
1103 
1104 static const struct file_operations spu_loadavg_fops = {
1105 	.open		= spu_loadavg_open,
1106 	.read		= seq_read,
1107 	.llseek		= seq_lseek,
1108 	.release	= single_release,
1109 };
1110 
1111 int __init spu_sched_init(void)
1112 {
1113 	struct proc_dir_entry *entry;
1114 	int err = -ENOMEM, i;
1115 
1116 	spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
1117 	if (!spu_prio)
1118 		goto out;
1119 
1120 	for (i = 0; i < MAX_PRIO; i++) {
1121 		INIT_LIST_HEAD(&spu_prio->runq[i]);
1122 		__clear_bit(i, spu_prio->bitmap);
1123 	}
1124 	spin_lock_init(&spu_prio->runq_lock);
1125 
1126 	setup_timer(&spusched_timer, spusched_wake, 0);
1127 	setup_timer(&spuloadavg_timer, spuloadavg_wake, 0);
1128 
1129 	spusched_task = kthread_run(spusched_thread, NULL, "spusched");
1130 	if (IS_ERR(spusched_task)) {
1131 		err = PTR_ERR(spusched_task);
1132 		goto out_free_spu_prio;
1133 	}
1134 
1135 	mod_timer(&spuloadavg_timer, 0);
1136 
1137 	entry = proc_create("spu_loadavg", 0, NULL, &spu_loadavg_fops);
1138 	if (!entry)
1139 		goto out_stop_kthread;
1140 
1141 	pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n",
1142 			SPUSCHED_TICK, MIN_SPU_TIMESLICE, DEF_SPU_TIMESLICE);
1143 	return 0;
1144 
1145  out_stop_kthread:
1146 	kthread_stop(spusched_task);
1147  out_free_spu_prio:
1148 	kfree(spu_prio);
1149  out:
1150 	return err;
1151 }
1152 
1153 void spu_sched_exit(void)
1154 {
1155 	struct spu *spu;
1156 	int node;
1157 
1158 	remove_proc_entry("spu_loadavg", NULL);
1159 
1160 	del_timer_sync(&spusched_timer);
1161 	del_timer_sync(&spuloadavg_timer);
1162 	kthread_stop(spusched_task);
1163 
1164 	for (node = 0; node < MAX_NUMNODES; node++) {
1165 		mutex_lock(&cbe_spu_info[node].list_mutex);
1166 		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list)
1167 			if (spu->alloc_state != SPU_FREE)
1168 				spu->alloc_state = SPU_FREE;
1169 		mutex_unlock(&cbe_spu_info[node].list_mutex);
1170 	}
1171 	kfree(spu_prio);
1172 }
1173