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