]>
Commit | Line | Data |
---|---|---|
8b3d6663 AB |
1 | /* sched.c - SPU scheduler. |
2 | * | |
3 | * Copyright (C) IBM 2005 | |
4 | * Author: Mark Nutter <mnutter@us.ibm.com> | |
5 | * | |
a68cf983 | 6 | * 2006-03-31 NUMA domains added. |
8b3d6663 AB |
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 | ||
3b3d22cb AB |
23 | #undef DEBUG |
24 | ||
8b3d6663 AB |
25 | #include <linux/module.h> |
26 | #include <linux/errno.h> | |
27 | #include <linux/sched.h> | |
28 | #include <linux/kernel.h> | |
29 | #include <linux/mm.h> | |
30 | #include <linux/completion.h> | |
31 | #include <linux/vmalloc.h> | |
32 | #include <linux/smp.h> | |
8b3d6663 AB |
33 | #include <linux/stddef.h> |
34 | #include <linux/unistd.h> | |
a68cf983 MN |
35 | #include <linux/numa.h> |
36 | #include <linux/mutex.h> | |
86767277 | 37 | #include <linux/notifier.h> |
37901802 | 38 | #include <linux/kthread.h> |
65de66f0 CH |
39 | #include <linux/pid_namespace.h> |
40 | #include <linux/proc_fs.h> | |
41 | #include <linux/seq_file.h> | |
8b3d6663 AB |
42 | |
43 | #include <asm/io.h> | |
44 | #include <asm/mmu_context.h> | |
45 | #include <asm/spu.h> | |
46 | #include <asm/spu_csa.h> | |
a91942ae | 47 | #include <asm/spu_priv1.h> |
8b3d6663 AB |
48 | #include "spufs.h" |
49 | ||
8b3d6663 | 50 | struct spu_prio_array { |
72cb3608 | 51 | DECLARE_BITMAP(bitmap, MAX_PRIO); |
079cdb61 CH |
52 | struct list_head runq[MAX_PRIO]; |
53 | spinlock_t runq_lock; | |
65de66f0 | 54 | int nr_waiting; |
8b3d6663 AB |
55 | }; |
56 | ||
65de66f0 | 57 | static unsigned long spu_avenrun[3]; |
a68cf983 | 58 | static struct spu_prio_array *spu_prio; |
37901802 CH |
59 | static struct task_struct *spusched_task; |
60 | static struct timer_list spusched_timer; | |
8b3d6663 | 61 | |
fe443ef2 CH |
62 | /* |
63 | * Priority of a normal, non-rt, non-niced'd process (aka nice level 0). | |
64 | */ | |
65 | #define NORMAL_PRIO 120 | |
66 | ||
67 | /* | |
68 | * Frequency of the spu scheduler tick. By default we do one SPU scheduler | |
69 | * tick for every 10 CPU scheduler ticks. | |
70 | */ | |
71 | #define SPUSCHED_TICK (10) | |
72 | ||
73 | /* | |
74 | * These are the 'tuning knobs' of the scheduler: | |
75 | * | |
60e24239 JK |
76 | * Minimum timeslice is 5 msecs (or 1 spu scheduler tick, whichever is |
77 | * larger), default timeslice is 100 msecs, maximum timeslice is 800 msecs. | |
fe443ef2 | 78 | */ |
60e24239 JK |
79 | #define MIN_SPU_TIMESLICE max(5 * HZ / (1000 * SPUSCHED_TICK), 1) |
80 | #define DEF_SPU_TIMESLICE (100 * HZ / (1000 * SPUSCHED_TICK)) | |
fe443ef2 CH |
81 | |
82 | #define MAX_USER_PRIO (MAX_PRIO - MAX_RT_PRIO) | |
83 | #define SCALE_PRIO(x, prio) \ | |
84 | max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_SPU_TIMESLICE) | |
85 | ||
86 | /* | |
87 | * scale user-nice values [ -20 ... 0 ... 19 ] to time slice values: | |
88 | * [800ms ... 100ms ... 5ms] | |
89 | * | |
90 | * The higher a thread's priority, the bigger timeslices | |
91 | * it gets during one round of execution. But even the lowest | |
92 | * priority thread gets MIN_TIMESLICE worth of execution time. | |
93 | */ | |
94 | void spu_set_timeslice(struct spu_context *ctx) | |
95 | { | |
96 | if (ctx->prio < NORMAL_PRIO) | |
97 | ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE * 4, ctx->prio); | |
98 | else | |
99 | ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE, ctx->prio); | |
100 | } | |
101 | ||
2cf2b3b4 CH |
102 | /* |
103 | * Update scheduling information from the owning thread. | |
104 | */ | |
105 | void __spu_update_sched_info(struct spu_context *ctx) | |
106 | { | |
91569531 LB |
107 | /* |
108 | * assert that the context is not on the runqueue, so it is safe | |
109 | * to change its scheduling parameters. | |
110 | */ | |
111 | BUG_ON(!list_empty(&ctx->rq)); | |
112 | ||
476273ad | 113 | /* |
9b1d21f8 JMV |
114 | * 32-Bit assignments are atomic on powerpc, and we don't care about |
115 | * memory ordering here because retrieving the controlling thread is | |
116 | * per definition racy. | |
476273ad CH |
117 | */ |
118 | ctx->tid = current->pid; | |
119 | ||
2cf2b3b4 CH |
120 | /* |
121 | * We do our own priority calculations, so we normally want | |
9b1d21f8 | 122 | * ->static_prio to start with. Unfortunately this field |
2cf2b3b4 CH |
123 | * contains junk for threads with a realtime scheduling |
124 | * policy so we have to look at ->prio in this case. | |
125 | */ | |
126 | if (rt_prio(current->prio)) | |
127 | ctx->prio = current->prio; | |
128 | else | |
129 | ctx->prio = current->static_prio; | |
130 | ctx->policy = current->policy; | |
ea1ae594 CH |
131 | |
132 | /* | |
91569531 LB |
133 | * TO DO: the context may be loaded, so we may need to activate |
134 | * it again on a different node. But it shouldn't hurt anything | |
135 | * to update its parameters, because we know that the scheduler | |
136 | * is not actively looking at this field, since it is not on the | |
137 | * runqueue. The context will be rescheduled on the proper node | |
138 | * if it is timesliced or preempted. | |
ea1ae594 | 139 | */ |
ea1ae594 | 140 | ctx->cpus_allowed = current->cpus_allowed; |
2cf2b3b4 CH |
141 | } |
142 | ||
143 | void spu_update_sched_info(struct spu_context *ctx) | |
144 | { | |
91569531 | 145 | int node; |
2cf2b3b4 | 146 | |
91569531 LB |
147 | if (ctx->state == SPU_STATE_RUNNABLE) { |
148 | node = ctx->spu->node; | |
149 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
150 | __spu_update_sched_info(ctx); | |
151 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
152 | } else { | |
153 | __spu_update_sched_info(ctx); | |
154 | } | |
2cf2b3b4 CH |
155 | } |
156 | ||
ea1ae594 | 157 | static int __node_allowed(struct spu_context *ctx, int node) |
8b3d6663 | 158 | { |
ea1ae594 CH |
159 | if (nr_cpus_node(node)) { |
160 | cpumask_t mask = node_to_cpumask(node); | |
8b3d6663 | 161 | |
ea1ae594 CH |
162 | if (cpus_intersects(mask, ctx->cpus_allowed)) |
163 | return 1; | |
164 | } | |
165 | ||
166 | return 0; | |
167 | } | |
168 | ||
169 | static int node_allowed(struct spu_context *ctx, int node) | |
170 | { | |
171 | int rval; | |
172 | ||
173 | spin_lock(&spu_prio->runq_lock); | |
174 | rval = __node_allowed(ctx, node); | |
175 | spin_unlock(&spu_prio->runq_lock); | |
176 | ||
177 | return rval; | |
8b3d6663 AB |
178 | } |
179 | ||
86767277 AB |
180 | static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier); |
181 | ||
36aaccc1 | 182 | void spu_switch_notify(struct spu *spu, struct spu_context *ctx) |
86767277 AB |
183 | { |
184 | blocking_notifier_call_chain(&spu_switch_notifier, | |
185 | ctx ? ctx->object_id : 0, spu); | |
186 | } | |
187 | ||
36aaccc1 BN |
188 | static void notify_spus_active(void) |
189 | { | |
190 | int node; | |
191 | ||
192 | /* | |
193 | * Wake up the active spu_contexts. | |
194 | * | |
195 | * When the awakened processes see their "notify_active" flag is set, | |
9b1d21f8 | 196 | * they will call spu_switch_notify(). |
36aaccc1 BN |
197 | */ |
198 | for_each_online_node(node) { | |
199 | struct spu *spu; | |
486acd48 CH |
200 | |
201 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
202 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { | |
203 | if (spu->alloc_state != SPU_FREE) { | |
204 | struct spu_context *ctx = spu->ctx; | |
205 | set_bit(SPU_SCHED_NOTIFY_ACTIVE, | |
206 | &ctx->sched_flags); | |
207 | mb(); | |
208 | wake_up_all(&ctx->stop_wq); | |
209 | } | |
36aaccc1 | 210 | } |
486acd48 | 211 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
36aaccc1 BN |
212 | } |
213 | } | |
214 | ||
86767277 AB |
215 | int spu_switch_event_register(struct notifier_block * n) |
216 | { | |
36aaccc1 BN |
217 | int ret; |
218 | ret = blocking_notifier_chain_register(&spu_switch_notifier, n); | |
219 | if (!ret) | |
220 | notify_spus_active(); | |
221 | return ret; | |
86767277 | 222 | } |
36aaccc1 | 223 | EXPORT_SYMBOL_GPL(spu_switch_event_register); |
86767277 AB |
224 | |
225 | int spu_switch_event_unregister(struct notifier_block * n) | |
226 | { | |
227 | return blocking_notifier_chain_unregister(&spu_switch_notifier, n); | |
228 | } | |
36aaccc1 | 229 | EXPORT_SYMBOL_GPL(spu_switch_event_unregister); |
86767277 | 230 | |
202557d2 CH |
231 | /** |
232 | * spu_bind_context - bind spu context to physical spu | |
233 | * @spu: physical spu to bind to | |
234 | * @ctx: context to bind | |
235 | */ | |
236 | static void spu_bind_context(struct spu *spu, struct spu_context *ctx) | |
8b3d6663 | 237 | { |
a68cf983 MN |
238 | pr_debug("%s: pid=%d SPU=%d NODE=%d\n", __FUNCTION__, current->pid, |
239 | spu->number, spu->node); | |
27ec41d3 | 240 | spuctx_switch_state(ctx, SPU_UTIL_SYSTEM); |
e9f8a0b6 | 241 | |
aa6d5b20 AB |
242 | if (ctx->flags & SPU_CREATE_NOSCHED) |
243 | atomic_inc(&cbe_spu_info[spu->node].reserved_spus); | |
244 | ||
e9f8a0b6 CH |
245 | ctx->stats.slb_flt_base = spu->stats.slb_flt; |
246 | ctx->stats.class2_intr_base = spu->stats.class2_intr; | |
247 | ||
8b3d6663 AB |
248 | spu->ctx = ctx; |
249 | spu->flags = 0; | |
250 | ctx->spu = spu; | |
251 | ctx->ops = &spu_hw_ops; | |
252 | spu->pid = current->pid; | |
1474855d | 253 | spu->tgid = current->tgid; |
94b2a439 | 254 | spu_associate_mm(spu, ctx->owner); |
8b3d6663 AB |
255 | spu->ibox_callback = spufs_ibox_callback; |
256 | spu->wbox_callback = spufs_wbox_callback; | |
5110459f | 257 | spu->stop_callback = spufs_stop_callback; |
a33a7d73 | 258 | spu->mfc_callback = spufs_mfc_callback; |
8b3d6663 | 259 | mb(); |
5110459f | 260 | spu_unmap_mappings(ctx); |
8b3d6663 | 261 | spu_restore(&ctx->csa, spu); |
2a911f0b | 262 | spu->timestamp = jiffies; |
a68cf983 | 263 | spu_cpu_affinity_set(spu, raw_smp_processor_id()); |
86767277 | 264 | spu_switch_notify(spu, ctx); |
81998baf | 265 | ctx->state = SPU_STATE_RUNNABLE; |
27ec41d3 AD |
266 | |
267 | spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED); | |
8b3d6663 AB |
268 | } |
269 | ||
c5fc8d2a | 270 | /* |
486acd48 | 271 | * Must be used with the list_mutex held. |
c5fc8d2a AB |
272 | */ |
273 | static inline int sched_spu(struct spu *spu) | |
274 | { | |
486acd48 CH |
275 | BUG_ON(!mutex_is_locked(&cbe_spu_info[spu->node].list_mutex)); |
276 | ||
c5fc8d2a AB |
277 | return (!spu->ctx || !(spu->ctx->flags & SPU_CREATE_NOSCHED)); |
278 | } | |
279 | ||
280 | static void aff_merge_remaining_ctxs(struct spu_gang *gang) | |
281 | { | |
282 | struct spu_context *ctx; | |
283 | ||
284 | list_for_each_entry(ctx, &gang->aff_list_head, aff_list) { | |
285 | if (list_empty(&ctx->aff_list)) | |
286 | list_add(&ctx->aff_list, &gang->aff_list_head); | |
287 | } | |
288 | gang->aff_flags |= AFF_MERGED; | |
289 | } | |
290 | ||
291 | static void aff_set_offsets(struct spu_gang *gang) | |
292 | { | |
293 | struct spu_context *ctx; | |
294 | int offset; | |
295 | ||
296 | offset = -1; | |
297 | list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list, | |
298 | aff_list) { | |
299 | if (&ctx->aff_list == &gang->aff_list_head) | |
300 | break; | |
301 | ctx->aff_offset = offset--; | |
302 | } | |
303 | ||
304 | offset = 0; | |
305 | list_for_each_entry(ctx, gang->aff_ref_ctx->aff_list.prev, aff_list) { | |
306 | if (&ctx->aff_list == &gang->aff_list_head) | |
307 | break; | |
308 | ctx->aff_offset = offset++; | |
309 | } | |
310 | ||
311 | gang->aff_flags |= AFF_OFFSETS_SET; | |
312 | } | |
313 | ||
314 | static struct spu *aff_ref_location(struct spu_context *ctx, int mem_aff, | |
315 | int group_size, int lowest_offset) | |
316 | { | |
317 | struct spu *spu; | |
318 | int node, n; | |
319 | ||
320 | /* | |
321 | * TODO: A better algorithm could be used to find a good spu to be | |
322 | * used as reference location for the ctxs chain. | |
323 | */ | |
324 | node = cpu_to_node(raw_smp_processor_id()); | |
325 | for (n = 0; n < MAX_NUMNODES; n++, node++) { | |
326 | node = (node < MAX_NUMNODES) ? node : 0; | |
327 | if (!node_allowed(ctx, node)) | |
328 | continue; | |
486acd48 | 329 | mutex_lock(&cbe_spu_info[node].list_mutex); |
c5fc8d2a AB |
330 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { |
331 | if ((!mem_aff || spu->has_mem_affinity) && | |
486acd48 CH |
332 | sched_spu(spu)) { |
333 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
c5fc8d2a | 334 | return spu; |
486acd48 | 335 | } |
c5fc8d2a | 336 | } |
486acd48 | 337 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
c5fc8d2a AB |
338 | } |
339 | return NULL; | |
340 | } | |
341 | ||
342 | static void aff_set_ref_point_location(struct spu_gang *gang) | |
343 | { | |
344 | int mem_aff, gs, lowest_offset; | |
345 | struct spu_context *ctx; | |
346 | struct spu *tmp; | |
347 | ||
348 | mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM; | |
349 | lowest_offset = 0; | |
350 | gs = 0; | |
351 | ||
352 | list_for_each_entry(tmp, &gang->aff_list_head, aff_list) | |
353 | gs++; | |
354 | ||
355 | list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list, | |
356 | aff_list) { | |
357 | if (&ctx->aff_list == &gang->aff_list_head) | |
358 | break; | |
359 | lowest_offset = ctx->aff_offset; | |
360 | } | |
361 | ||
683e3ab2 AD |
362 | gang->aff_ref_spu = aff_ref_location(gang->aff_ref_ctx, mem_aff, gs, |
363 | lowest_offset); | |
c5fc8d2a AB |
364 | } |
365 | ||
486acd48 | 366 | static struct spu *ctx_location(struct spu *ref, int offset, int node) |
c5fc8d2a AB |
367 | { |
368 | struct spu *spu; | |
369 | ||
370 | spu = NULL; | |
371 | if (offset >= 0) { | |
372 | list_for_each_entry(spu, ref->aff_list.prev, aff_list) { | |
486acd48 | 373 | BUG_ON(spu->node != node); |
c5fc8d2a AB |
374 | if (offset == 0) |
375 | break; | |
376 | if (sched_spu(spu)) | |
377 | offset--; | |
378 | } | |
379 | } else { | |
380 | list_for_each_entry_reverse(spu, ref->aff_list.next, aff_list) { | |
486acd48 | 381 | BUG_ON(spu->node != node); |
c5fc8d2a AB |
382 | if (offset == 0) |
383 | break; | |
384 | if (sched_spu(spu)) | |
385 | offset++; | |
386 | } | |
387 | } | |
486acd48 | 388 | |
c5fc8d2a AB |
389 | return spu; |
390 | } | |
391 | ||
392 | /* | |
393 | * affinity_check is called each time a context is going to be scheduled. | |
394 | * It returns the spu ptr on which the context must run. | |
395 | */ | |
486acd48 | 396 | static int has_affinity(struct spu_context *ctx) |
c5fc8d2a | 397 | { |
486acd48 | 398 | struct spu_gang *gang = ctx->gang; |
c5fc8d2a AB |
399 | |
400 | if (list_empty(&ctx->aff_list)) | |
486acd48 CH |
401 | return 0; |
402 | ||
c5fc8d2a AB |
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 | } | |
486acd48 CH |
410 | |
411 | return gang->aff_ref_spu != NULL; | |
c5fc8d2a AB |
412 | } |
413 | ||
202557d2 CH |
414 | /** |
415 | * spu_unbind_context - unbind spu context from physical spu | |
416 | * @spu: physical spu to unbind from | |
417 | * @ctx: context to unbind | |
202557d2 | 418 | */ |
678b2ff1 | 419 | static void spu_unbind_context(struct spu *spu, struct spu_context *ctx) |
8b3d6663 | 420 | { |
a68cf983 MN |
421 | pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__, |
422 | spu->pid, spu->number, spu->node); | |
27ec41d3 | 423 | spuctx_switch_state(ctx, SPU_UTIL_SYSTEM); |
fe2f896d | 424 | |
aa6d5b20 AB |
425 | if (spu->ctx->flags & SPU_CREATE_NOSCHED) |
426 | atomic_dec(&cbe_spu_info[spu->node].reserved_spus); | |
36ddbb13 AD |
427 | |
428 | if (ctx->gang){ | |
429 | mutex_lock(&ctx->gang->aff_mutex); | |
430 | if (has_affinity(ctx)) { | |
431 | if (atomic_dec_and_test(&ctx->gang->aff_sched_count)) | |
432 | ctx->gang->aff_ref_spu = NULL; | |
433 | } | |
434 | mutex_unlock(&ctx->gang->aff_mutex); | |
435 | } | |
436 | ||
86767277 | 437 | spu_switch_notify(spu, NULL); |
5110459f | 438 | spu_unmap_mappings(ctx); |
8b3d6663 | 439 | spu_save(&ctx->csa, spu); |
2a911f0b | 440 | spu->timestamp = jiffies; |
8b3d6663 AB |
441 | ctx->state = SPU_STATE_SAVED; |
442 | spu->ibox_callback = NULL; | |
443 | spu->wbox_callback = NULL; | |
5110459f | 444 | spu->stop_callback = NULL; |
a33a7d73 | 445 | spu->mfc_callback = NULL; |
94b2a439 | 446 | spu_associate_mm(spu, NULL); |
8b3d6663 | 447 | spu->pid = 0; |
1474855d | 448 | spu->tgid = 0; |
8b3d6663 | 449 | ctx->ops = &spu_backing_ops; |
2a911f0b | 450 | spu->flags = 0; |
8b3d6663 | 451 | spu->ctx = NULL; |
e9f8a0b6 CH |
452 | |
453 | ctx->stats.slb_flt += | |
454 | (spu->stats.slb_flt - ctx->stats.slb_flt_base); | |
455 | ctx->stats.class2_intr += | |
456 | (spu->stats.class2_intr - ctx->stats.class2_intr_base); | |
27ec41d3 AD |
457 | |
458 | /* This maps the underlying spu state to idle */ | |
459 | spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED); | |
460 | ctx->spu = NULL; | |
8b3d6663 AB |
461 | } |
462 | ||
079cdb61 CH |
463 | /** |
464 | * spu_add_to_rq - add a context to the runqueue | |
465 | * @ctx: context to add | |
466 | */ | |
4e0f4ed0 | 467 | static void __spu_add_to_rq(struct spu_context *ctx) |
8b3d6663 | 468 | { |
27449971 CH |
469 | /* |
470 | * Unfortunately this code path can be called from multiple threads | |
471 | * on behalf of a single context due to the way the problem state | |
472 | * mmap support works. | |
473 | * | |
474 | * Fortunately we need to wake up all these threads at the same time | |
475 | * and can simply skip the runqueue addition for every but the first | |
476 | * thread getting into this codepath. | |
477 | * | |
478 | * It's still quite hacky, and long-term we should proxy all other | |
479 | * threads through the owner thread so that spu_run is in control | |
480 | * of all the scheduling activity for a given context. | |
481 | */ | |
482 | if (list_empty(&ctx->rq)) { | |
483 | list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]); | |
484 | set_bit(ctx->prio, spu_prio->bitmap); | |
485 | if (!spu_prio->nr_waiting++) | |
486 | __mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK); | |
487 | } | |
2a911f0b | 488 | } |
5110459f | 489 | |
4e0f4ed0 | 490 | static void __spu_del_from_rq(struct spu_context *ctx) |
a475c2f4 | 491 | { |
4e0f4ed0 LB |
492 | int prio = ctx->prio; |
493 | ||
65de66f0 | 494 | if (!list_empty(&ctx->rq)) { |
c77239b8 CH |
495 | if (!--spu_prio->nr_waiting) |
496 | del_timer(&spusched_timer); | |
a475c2f4 | 497 | list_del_init(&ctx->rq); |
c77239b8 CH |
498 | |
499 | if (list_empty(&spu_prio->runq[prio])) | |
500 | clear_bit(prio, spu_prio->bitmap); | |
65de66f0 | 501 | } |
079cdb61 | 502 | } |
a68cf983 | 503 | |
079cdb61 | 504 | static void spu_prio_wait(struct spu_context *ctx) |
8b3d6663 | 505 | { |
a68cf983 | 506 | DEFINE_WAIT(wait); |
8b3d6663 | 507 | |
4e0f4ed0 | 508 | spin_lock(&spu_prio->runq_lock); |
079cdb61 | 509 | prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE); |
a68cf983 | 510 | if (!signal_pending(current)) { |
4e0f4ed0 LB |
511 | __spu_add_to_rq(ctx); |
512 | spin_unlock(&spu_prio->runq_lock); | |
650f8b02 | 513 | mutex_unlock(&ctx->state_mutex); |
a68cf983 | 514 | schedule(); |
650f8b02 | 515 | mutex_lock(&ctx->state_mutex); |
4e0f4ed0 LB |
516 | spin_lock(&spu_prio->runq_lock); |
517 | __spu_del_from_rq(ctx); | |
8b3d6663 | 518 | } |
4e0f4ed0 | 519 | spin_unlock(&spu_prio->runq_lock); |
079cdb61 CH |
520 | __set_current_state(TASK_RUNNING); |
521 | remove_wait_queue(&ctx->stop_wq, &wait); | |
8b3d6663 AB |
522 | } |
523 | ||
079cdb61 | 524 | static struct spu *spu_get_idle(struct spu_context *ctx) |
a68cf983 | 525 | { |
36ddbb13 | 526 | struct spu *spu, *aff_ref_spu; |
486acd48 CH |
527 | int node, n; |
528 | ||
36ddbb13 AD |
529 | if (ctx->gang) { |
530 | mutex_lock(&ctx->gang->aff_mutex); | |
531 | if (has_affinity(ctx)) { | |
532 | aff_ref_spu = ctx->gang->aff_ref_spu; | |
533 | atomic_inc(&ctx->gang->aff_sched_count); | |
534 | mutex_unlock(&ctx->gang->aff_mutex); | |
535 | node = aff_ref_spu->node; | |
536 | ||
537 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
538 | spu = ctx_location(aff_ref_spu, ctx->aff_offset, node); | |
539 | if (spu && spu->alloc_state == SPU_FREE) | |
540 | goto found; | |
541 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
a68cf983 | 542 | |
36ddbb13 AD |
543 | mutex_lock(&ctx->gang->aff_mutex); |
544 | if (atomic_dec_and_test(&ctx->gang->aff_sched_count)) | |
545 | ctx->gang->aff_ref_spu = NULL; | |
546 | mutex_unlock(&ctx->gang->aff_mutex); | |
cbc23d3e | 547 | |
36ddbb13 AD |
548 | return NULL; |
549 | } | |
550 | mutex_unlock(&ctx->gang->aff_mutex); | |
551 | } | |
486acd48 | 552 | node = cpu_to_node(raw_smp_processor_id()); |
a68cf983 MN |
553 | for (n = 0; n < MAX_NUMNODES; n++, node++) { |
554 | node = (node < MAX_NUMNODES) ? node : 0; | |
ea1ae594 | 555 | if (!node_allowed(ctx, node)) |
a68cf983 | 556 | continue; |
486acd48 CH |
557 | |
558 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
559 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { | |
560 | if (spu->alloc_state == SPU_FREE) | |
561 | goto found; | |
562 | } | |
563 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
a68cf983 | 564 | } |
486acd48 CH |
565 | |
566 | return NULL; | |
567 | ||
568 | found: | |
569 | spu->alloc_state = SPU_USED; | |
570 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
571 | pr_debug("Got SPU %d %d\n", spu->number, spu->node); | |
572 | spu_init_channels(spu); | |
a68cf983 MN |
573 | return spu; |
574 | } | |
8b3d6663 | 575 | |
52f04fcf CH |
576 | /** |
577 | * find_victim - find a lower priority context to preempt | |
578 | * @ctx: canidate context for running | |
579 | * | |
580 | * Returns the freed physical spu to run the new context on. | |
581 | */ | |
582 | static struct spu *find_victim(struct spu_context *ctx) | |
583 | { | |
584 | struct spu_context *victim = NULL; | |
585 | struct spu *spu; | |
586 | int node, n; | |
587 | ||
588 | /* | |
589 | * Look for a possible preemption candidate on the local node first. | |
590 | * If there is no candidate look at the other nodes. This isn't | |
9b1d21f8 | 591 | * exactly fair, but so far the whole spu scheduler tries to keep |
52f04fcf CH |
592 | * a strong node affinity. We might want to fine-tune this in |
593 | * the future. | |
594 | */ | |
595 | restart: | |
596 | node = cpu_to_node(raw_smp_processor_id()); | |
597 | for (n = 0; n < MAX_NUMNODES; n++, node++) { | |
598 | node = (node < MAX_NUMNODES) ? node : 0; | |
ea1ae594 | 599 | if (!node_allowed(ctx, node)) |
52f04fcf CH |
600 | continue; |
601 | ||
486acd48 CH |
602 | mutex_lock(&cbe_spu_info[node].list_mutex); |
603 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { | |
52f04fcf CH |
604 | struct spu_context *tmp = spu->ctx; |
605 | ||
c0e7b4aa | 606 | if (tmp && tmp->prio > ctx->prio && |
fe443ef2 | 607 | (!victim || tmp->prio > victim->prio)) |
52f04fcf CH |
608 | victim = spu->ctx; |
609 | } | |
486acd48 | 610 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
52f04fcf CH |
611 | |
612 | if (victim) { | |
613 | /* | |
614 | * This nests ctx->state_mutex, but we always lock | |
615 | * higher priority contexts before lower priority | |
616 | * ones, so this is safe until we introduce | |
617 | * priority inheritance schemes. | |
91569531 LB |
618 | * |
619 | * XXX if the highest priority context is locked, | |
620 | * this can loop a long time. Might be better to | |
621 | * look at another context or give up after X retries. | |
52f04fcf CH |
622 | */ |
623 | if (!mutex_trylock(&victim->state_mutex)) { | |
624 | victim = NULL; | |
625 | goto restart; | |
626 | } | |
627 | ||
628 | spu = victim->spu; | |
b192541b | 629 | if (!spu || victim->prio <= ctx->prio) { |
52f04fcf CH |
630 | /* |
631 | * This race can happen because we've dropped | |
b192541b | 632 | * the active list mutex. Not a problem, just |
52f04fcf CH |
633 | * restart the search. |
634 | */ | |
635 | mutex_unlock(&victim->state_mutex); | |
636 | victim = NULL; | |
637 | goto restart; | |
638 | } | |
486acd48 CH |
639 | |
640 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
641 | cbe_spu_info[node].nr_active--; | |
c0e7b4aa | 642 | spu_unbind_context(spu, victim); |
486acd48 CH |
643 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
644 | ||
e9f8a0b6 | 645 | victim->stats.invol_ctx_switch++; |
fe2f896d | 646 | spu->stats.invol_ctx_switch++; |
52f04fcf | 647 | mutex_unlock(&victim->state_mutex); |
e097b513 CH |
648 | /* |
649 | * We need to break out of the wait loop in spu_run | |
650 | * manually to ensure this context gets put on the | |
651 | * runqueue again ASAP. | |
652 | */ | |
653 | wake_up(&victim->stop_wq); | |
52f04fcf CH |
654 | return spu; |
655 | } | |
656 | } | |
657 | ||
658 | return NULL; | |
659 | } | |
660 | ||
079cdb61 CH |
661 | /** |
662 | * spu_activate - find a free spu for a context and execute it | |
663 | * @ctx: spu context to schedule | |
664 | * @flags: flags (currently ignored) | |
665 | * | |
08873095 | 666 | * Tries to find a free spu to run @ctx. If no free spu is available |
079cdb61 CH |
667 | * add the context to the runqueue so it gets woken up once an spu |
668 | * is available. | |
669 | */ | |
26bec673 | 670 | int spu_activate(struct spu_context *ctx, unsigned long flags) |
8b3d6663 | 671 | { |
079cdb61 CH |
672 | do { |
673 | struct spu *spu; | |
674 | ||
27449971 CH |
675 | /* |
676 | * If there are multiple threads waiting for a single context | |
677 | * only one actually binds the context while the others will | |
678 | * only be able to acquire the state_mutex once the context | |
679 | * already is in runnable state. | |
680 | */ | |
681 | if (ctx->spu) | |
682 | return 0; | |
683 | ||
079cdb61 | 684 | spu = spu_get_idle(ctx); |
52f04fcf CH |
685 | /* |
686 | * If this is a realtime thread we try to get it running by | |
687 | * preempting a lower priority thread. | |
688 | */ | |
fe443ef2 | 689 | if (!spu && rt_prio(ctx->prio)) |
52f04fcf | 690 | spu = find_victim(ctx); |
079cdb61 | 691 | if (spu) { |
486acd48 CH |
692 | int node = spu->node; |
693 | ||
694 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
202557d2 | 695 | spu_bind_context(spu, ctx); |
486acd48 CH |
696 | cbe_spu_info[node].nr_active++; |
697 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
33bfd7a7 | 698 | wake_up_all(&ctx->run_wq); |
079cdb61 | 699 | return 0; |
a68cf983 | 700 | } |
079cdb61 | 701 | |
50b520d4 | 702 | spu_prio_wait(ctx); |
079cdb61 CH |
703 | } while (!signal_pending(current)); |
704 | ||
705 | return -ERESTARTSYS; | |
8b3d6663 AB |
706 | } |
707 | ||
bb5db29a CH |
708 | /** |
709 | * grab_runnable_context - try to find a runnable context | |
710 | * | |
711 | * Remove the highest priority context on the runqueue and return it | |
712 | * to the caller. Returns %NULL if no runnable context was found. | |
713 | */ | |
ea1ae594 | 714 | static struct spu_context *grab_runnable_context(int prio, int node) |
bb5db29a | 715 | { |
ea1ae594 | 716 | struct spu_context *ctx; |
bb5db29a CH |
717 | int best; |
718 | ||
719 | spin_lock(&spu_prio->runq_lock); | |
7e90b749 | 720 | best = find_first_bit(spu_prio->bitmap, prio); |
ea1ae594 | 721 | while (best < prio) { |
bb5db29a CH |
722 | struct list_head *rq = &spu_prio->runq[best]; |
723 | ||
ea1ae594 CH |
724 | list_for_each_entry(ctx, rq, rq) { |
725 | /* XXX(hch): check for affinity here aswell */ | |
726 | if (__node_allowed(ctx, node)) { | |
727 | __spu_del_from_rq(ctx); | |
728 | goto found; | |
729 | } | |
730 | } | |
731 | best++; | |
bb5db29a | 732 | } |
ea1ae594 CH |
733 | ctx = NULL; |
734 | found: | |
bb5db29a | 735 | spin_unlock(&spu_prio->runq_lock); |
bb5db29a CH |
736 | return ctx; |
737 | } | |
738 | ||
739 | static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio) | |
740 | { | |
741 | struct spu *spu = ctx->spu; | |
742 | struct spu_context *new = NULL; | |
743 | ||
744 | if (spu) { | |
ea1ae594 | 745 | new = grab_runnable_context(max_prio, spu->node); |
bb5db29a | 746 | if (new || force) { |
486acd48 CH |
747 | int node = spu->node; |
748 | ||
749 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
bb5db29a | 750 | spu_unbind_context(spu, ctx); |
486acd48 CH |
751 | spu->alloc_state = SPU_FREE; |
752 | cbe_spu_info[node].nr_active--; | |
753 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
754 | ||
e9f8a0b6 | 755 | ctx->stats.vol_ctx_switch++; |
fe2f896d | 756 | spu->stats.vol_ctx_switch++; |
486acd48 | 757 | |
bb5db29a CH |
758 | if (new) |
759 | wake_up(&new->stop_wq); | |
760 | } | |
761 | ||
762 | } | |
763 | ||
764 | return new != NULL; | |
765 | } | |
766 | ||
678b2ff1 CH |
767 | /** |
768 | * spu_deactivate - unbind a context from it's physical spu | |
769 | * @ctx: spu context to unbind | |
770 | * | |
771 | * Unbind @ctx from the physical spu it is running on and schedule | |
772 | * the highest priority context to run on the freed physical spu. | |
773 | */ | |
8b3d6663 AB |
774 | void spu_deactivate(struct spu_context *ctx) |
775 | { | |
bb5db29a | 776 | __spu_deactivate(ctx, 1, MAX_PRIO); |
8b3d6663 AB |
777 | } |
778 | ||
ae7b4c52 | 779 | /** |
1474855d | 780 | * spu_yield - yield a physical spu if others are waiting |
ae7b4c52 CH |
781 | * @ctx: spu context to yield |
782 | * | |
783 | * Check if there is a higher priority context waiting and if yes | |
784 | * unbind @ctx from the physical spu and schedule the highest | |
785 | * priority context to run on the freed physical spu instead. | |
786 | */ | |
8b3d6663 AB |
787 | void spu_yield(struct spu_context *ctx) |
788 | { | |
e5c0b9ec CH |
789 | if (!(ctx->flags & SPU_CREATE_NOSCHED)) { |
790 | mutex_lock(&ctx->state_mutex); | |
27ec41d3 | 791 | __spu_deactivate(ctx, 0, MAX_PRIO); |
e5c0b9ec CH |
792 | mutex_unlock(&ctx->state_mutex); |
793 | } | |
bb5db29a | 794 | } |
8b3d6663 | 795 | |
486acd48 | 796 | static noinline void spusched_tick(struct spu_context *ctx) |
bb5db29a | 797 | { |
df09cf3e CH |
798 | if (ctx->flags & SPU_CREATE_NOSCHED) |
799 | return; | |
800 | if (ctx->policy == SCHED_FIFO) | |
801 | return; | |
802 | ||
803 | if (--ctx->time_slice) | |
37901802 | 804 | return; |
bb5db29a CH |
805 | |
806 | /* | |
486acd48 | 807 | * Unfortunately list_mutex ranks outside of state_mutex, so |
37901802 CH |
808 | * we have to trylock here. If we fail give the context another |
809 | * tick and try again. | |
bb5db29a | 810 | */ |
37901802 | 811 | if (mutex_trylock(&ctx->state_mutex)) { |
7022543e | 812 | struct spu *spu = ctx->spu; |
ea1ae594 CH |
813 | struct spu_context *new; |
814 | ||
815 | new = grab_runnable_context(ctx->prio + 1, spu->node); | |
37901802 | 816 | if (new) { |
37901802 | 817 | spu_unbind_context(spu, ctx); |
e9f8a0b6 | 818 | ctx->stats.invol_ctx_switch++; |
fe2f896d | 819 | spu->stats.invol_ctx_switch++; |
486acd48 CH |
820 | spu->alloc_state = SPU_FREE; |
821 | cbe_spu_info[spu->node].nr_active--; | |
37901802 CH |
822 | wake_up(&new->stop_wq); |
823 | /* | |
824 | * We need to break out of the wait loop in | |
825 | * spu_run manually to ensure this context | |
826 | * gets put on the runqueue again ASAP. | |
827 | */ | |
828 | wake_up(&ctx->stop_wq); | |
829 | } | |
fe443ef2 | 830 | spu_set_timeslice(ctx); |
37901802 | 831 | mutex_unlock(&ctx->state_mutex); |
bb5db29a | 832 | } else { |
37901802 | 833 | ctx->time_slice++; |
8b3d6663 | 834 | } |
8b3d6663 AB |
835 | } |
836 | ||
65de66f0 CH |
837 | /** |
838 | * count_active_contexts - count nr of active tasks | |
839 | * | |
840 | * Return the number of tasks currently running or waiting to run. | |
841 | * | |
486acd48 | 842 | * Note that we don't take runq_lock / list_mutex here. Reading |
65de66f0 CH |
843 | * a single 32bit value is atomic on powerpc, and we don't care |
844 | * about memory ordering issues here. | |
845 | */ | |
846 | static unsigned long count_active_contexts(void) | |
847 | { | |
848 | int nr_active = 0, node; | |
849 | ||
850 | for (node = 0; node < MAX_NUMNODES; node++) | |
486acd48 | 851 | nr_active += cbe_spu_info[node].nr_active; |
65de66f0 CH |
852 | nr_active += spu_prio->nr_waiting; |
853 | ||
854 | return nr_active; | |
855 | } | |
856 | ||
857 | /** | |
858 | * spu_calc_load - given tick count, update the avenrun load estimates. | |
859 | * @tick: tick count | |
860 | * | |
861 | * No locking against reading these values from userspace, as for | |
862 | * the CPU loadavg code. | |
863 | */ | |
864 | static void spu_calc_load(unsigned long ticks) | |
865 | { | |
866 | unsigned long active_tasks; /* fixed-point */ | |
867 | static int count = LOAD_FREQ; | |
868 | ||
869 | count -= ticks; | |
870 | ||
871 | if (unlikely(count < 0)) { | |
872 | active_tasks = count_active_contexts() * FIXED_1; | |
873 | do { | |
874 | CALC_LOAD(spu_avenrun[0], EXP_1, active_tasks); | |
875 | CALC_LOAD(spu_avenrun[1], EXP_5, active_tasks); | |
876 | CALC_LOAD(spu_avenrun[2], EXP_15, active_tasks); | |
877 | count += LOAD_FREQ; | |
878 | } while (count < 0); | |
879 | } | |
880 | } | |
881 | ||
37901802 CH |
882 | static void spusched_wake(unsigned long data) |
883 | { | |
884 | mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK); | |
885 | wake_up_process(spusched_task); | |
65de66f0 | 886 | spu_calc_load(SPUSCHED_TICK); |
37901802 CH |
887 | } |
888 | ||
889 | static int spusched_thread(void *unused) | |
890 | { | |
486acd48 | 891 | struct spu *spu; |
37901802 CH |
892 | int node; |
893 | ||
37901802 CH |
894 | while (!kthread_should_stop()) { |
895 | set_current_state(TASK_INTERRUPTIBLE); | |
896 | schedule(); | |
897 | for (node = 0; node < MAX_NUMNODES; node++) { | |
486acd48 CH |
898 | mutex_lock(&cbe_spu_info[node].list_mutex); |
899 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) | |
900 | if (spu->ctx) | |
901 | spusched_tick(spu->ctx); | |
902 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
37901802 CH |
903 | } |
904 | } | |
905 | ||
37901802 CH |
906 | return 0; |
907 | } | |
908 | ||
7cd58e43 JK |
909 | void spuctx_switch_state(struct spu_context *ctx, |
910 | enum spu_utilization_state new_state) | |
911 | { | |
912 | unsigned long long curtime; | |
913 | signed long long delta; | |
914 | struct timespec ts; | |
915 | struct spu *spu; | |
916 | enum spu_utilization_state old_state; | |
917 | ||
918 | ktime_get_ts(&ts); | |
919 | curtime = timespec_to_ns(&ts); | |
920 | delta = curtime - ctx->stats.tstamp; | |
921 | ||
922 | WARN_ON(!mutex_is_locked(&ctx->state_mutex)); | |
923 | WARN_ON(delta < 0); | |
924 | ||
925 | spu = ctx->spu; | |
926 | old_state = ctx->stats.util_state; | |
927 | ctx->stats.util_state = new_state; | |
928 | ctx->stats.tstamp = curtime; | |
929 | ||
930 | /* | |
931 | * Update the physical SPU utilization statistics. | |
932 | */ | |
933 | if (spu) { | |
934 | ctx->stats.times[old_state] += delta; | |
935 | spu->stats.times[old_state] += delta; | |
936 | spu->stats.util_state = new_state; | |
937 | spu->stats.tstamp = curtime; | |
938 | } | |
939 | } | |
940 | ||
65de66f0 CH |
941 | #define LOAD_INT(x) ((x) >> FSHIFT) |
942 | #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100) | |
943 | ||
944 | static int show_spu_loadavg(struct seq_file *s, void *private) | |
945 | { | |
946 | int a, b, c; | |
947 | ||
948 | a = spu_avenrun[0] + (FIXED_1/200); | |
949 | b = spu_avenrun[1] + (FIXED_1/200); | |
950 | c = spu_avenrun[2] + (FIXED_1/200); | |
951 | ||
952 | /* | |
953 | * Note that last_pid doesn't really make much sense for the | |
9b1d21f8 | 954 | * SPU loadavg (it even seems very odd on the CPU side...), |
65de66f0 CH |
955 | * but we include it here to have a 100% compatible interface. |
956 | */ | |
957 | seq_printf(s, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n", | |
958 | LOAD_INT(a), LOAD_FRAC(a), | |
959 | LOAD_INT(b), LOAD_FRAC(b), | |
960 | LOAD_INT(c), LOAD_FRAC(c), | |
961 | count_active_contexts(), | |
962 | atomic_read(&nr_spu_contexts), | |
963 | current->nsproxy->pid_ns->last_pid); | |
964 | return 0; | |
965 | } | |
966 | ||
967 | static int spu_loadavg_open(struct inode *inode, struct file *file) | |
968 | { | |
969 | return single_open(file, show_spu_loadavg, NULL); | |
970 | } | |
971 | ||
972 | static const struct file_operations spu_loadavg_fops = { | |
973 | .open = spu_loadavg_open, | |
974 | .read = seq_read, | |
975 | .llseek = seq_lseek, | |
976 | .release = single_release, | |
977 | }; | |
978 | ||
8b3d6663 AB |
979 | int __init spu_sched_init(void) |
980 | { | |
65de66f0 CH |
981 | struct proc_dir_entry *entry; |
982 | int err = -ENOMEM, i; | |
8b3d6663 | 983 | |
a68cf983 | 984 | spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL); |
37901802 | 985 | if (!spu_prio) |
65de66f0 | 986 | goto out; |
37901802 | 987 | |
8b3d6663 | 988 | for (i = 0; i < MAX_PRIO; i++) { |
079cdb61 | 989 | INIT_LIST_HEAD(&spu_prio->runq[i]); |
a68cf983 | 990 | __clear_bit(i, spu_prio->bitmap); |
8b3d6663 | 991 | } |
079cdb61 | 992 | spin_lock_init(&spu_prio->runq_lock); |
37901802 | 993 | |
c77239b8 CH |
994 | setup_timer(&spusched_timer, spusched_wake, 0); |
995 | ||
37901802 CH |
996 | spusched_task = kthread_run(spusched_thread, NULL, "spusched"); |
997 | if (IS_ERR(spusched_task)) { | |
65de66f0 CH |
998 | err = PTR_ERR(spusched_task); |
999 | goto out_free_spu_prio; | |
37901802 | 1000 | } |
f3f59bec | 1001 | |
65de66f0 CH |
1002 | entry = create_proc_entry("spu_loadavg", 0, NULL); |
1003 | if (!entry) | |
1004 | goto out_stop_kthread; | |
1005 | entry->proc_fops = &spu_loadavg_fops; | |
1006 | ||
f3f59bec JK |
1007 | pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n", |
1008 | SPUSCHED_TICK, MIN_SPU_TIMESLICE, DEF_SPU_TIMESLICE); | |
8b3d6663 | 1009 | return 0; |
37901802 | 1010 | |
65de66f0 CH |
1011 | out_stop_kthread: |
1012 | kthread_stop(spusched_task); | |
1013 | out_free_spu_prio: | |
1014 | kfree(spu_prio); | |
1015 | out: | |
1016 | return err; | |
8b3d6663 AB |
1017 | } |
1018 | ||
d1450317 | 1019 | void spu_sched_exit(void) |
8b3d6663 | 1020 | { |
486acd48 | 1021 | struct spu *spu; |
a68cf983 MN |
1022 | int node; |
1023 | ||
65de66f0 CH |
1024 | remove_proc_entry("spu_loadavg", NULL); |
1025 | ||
c77239b8 | 1026 | del_timer_sync(&spusched_timer); |
37901802 CH |
1027 | kthread_stop(spusched_task); |
1028 | ||
a68cf983 | 1029 | for (node = 0; node < MAX_NUMNODES; node++) { |
486acd48 CH |
1030 | mutex_lock(&cbe_spu_info[node].list_mutex); |
1031 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) | |
1032 | if (spu->alloc_state != SPU_FREE) | |
1033 | spu->alloc_state = SPU_FREE; | |
1034 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
8b3d6663 | 1035 | } |
a68cf983 | 1036 | kfree(spu_prio); |
8b3d6663 | 1037 | } |