]>
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> | |
038200cf | 42 | #include <linux/marker.h> |
8b3d6663 AB |
43 | |
44 | #include <asm/io.h> | |
45 | #include <asm/mmu_context.h> | |
46 | #include <asm/spu.h> | |
47 | #include <asm/spu_csa.h> | |
a91942ae | 48 | #include <asm/spu_priv1.h> |
8b3d6663 AB |
49 | #include "spufs.h" |
50 | ||
8b3d6663 | 51 | struct spu_prio_array { |
72cb3608 | 52 | DECLARE_BITMAP(bitmap, MAX_PRIO); |
079cdb61 CH |
53 | struct list_head runq[MAX_PRIO]; |
54 | spinlock_t runq_lock; | |
65de66f0 | 55 | int nr_waiting; |
8b3d6663 AB |
56 | }; |
57 | ||
65de66f0 | 58 | static unsigned long spu_avenrun[3]; |
a68cf983 | 59 | static struct spu_prio_array *spu_prio; |
37901802 CH |
60 | static struct task_struct *spusched_task; |
61 | static struct timer_list spusched_timer; | |
90608a29 | 62 | static struct timer_list spuloadavg_timer; |
8b3d6663 | 63 | |
fe443ef2 CH |
64 | /* |
65 | * Priority of a normal, non-rt, non-niced'd process (aka nice level 0). | |
66 | */ | |
67 | #define NORMAL_PRIO 120 | |
68 | ||
69 | /* | |
70 | * Frequency of the spu scheduler tick. By default we do one SPU scheduler | |
71 | * tick for every 10 CPU scheduler ticks. | |
72 | */ | |
73 | #define SPUSCHED_TICK (10) | |
74 | ||
75 | /* | |
76 | * These are the 'tuning knobs' of the scheduler: | |
77 | * | |
60e24239 JK |
78 | * Minimum timeslice is 5 msecs (or 1 spu scheduler tick, whichever is |
79 | * larger), default timeslice is 100 msecs, maximum timeslice is 800 msecs. | |
fe443ef2 | 80 | */ |
60e24239 JK |
81 | #define MIN_SPU_TIMESLICE max(5 * HZ / (1000 * SPUSCHED_TICK), 1) |
82 | #define DEF_SPU_TIMESLICE (100 * HZ / (1000 * SPUSCHED_TICK)) | |
fe443ef2 CH |
83 | |
84 | #define MAX_USER_PRIO (MAX_PRIO - MAX_RT_PRIO) | |
85 | #define SCALE_PRIO(x, prio) \ | |
86 | max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_SPU_TIMESLICE) | |
87 | ||
88 | /* | |
89 | * scale user-nice values [ -20 ... 0 ... 19 ] to time slice values: | |
90 | * [800ms ... 100ms ... 5ms] | |
91 | * | |
92 | * The higher a thread's priority, the bigger timeslices | |
93 | * it gets during one round of execution. But even the lowest | |
94 | * priority thread gets MIN_TIMESLICE worth of execution time. | |
95 | */ | |
96 | void spu_set_timeslice(struct spu_context *ctx) | |
97 | { | |
98 | if (ctx->prio < NORMAL_PRIO) | |
99 | ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE * 4, ctx->prio); | |
100 | else | |
101 | ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE, ctx->prio); | |
102 | } | |
103 | ||
2cf2b3b4 CH |
104 | /* |
105 | * Update scheduling information from the owning thread. | |
106 | */ | |
107 | void __spu_update_sched_info(struct spu_context *ctx) | |
108 | { | |
91569531 LB |
109 | /* |
110 | * assert that the context is not on the runqueue, so it is safe | |
111 | * to change its scheduling parameters. | |
112 | */ | |
113 | BUG_ON(!list_empty(&ctx->rq)); | |
114 | ||
476273ad | 115 | /* |
9b1d21f8 JMV |
116 | * 32-Bit assignments are atomic on powerpc, and we don't care about |
117 | * memory ordering here because retrieving the controlling thread is | |
118 | * per definition racy. | |
476273ad CH |
119 | */ |
120 | ctx->tid = current->pid; | |
121 | ||
2cf2b3b4 CH |
122 | /* |
123 | * We do our own priority calculations, so we normally want | |
9b1d21f8 | 124 | * ->static_prio to start with. Unfortunately this field |
2cf2b3b4 CH |
125 | * contains junk for threads with a realtime scheduling |
126 | * policy so we have to look at ->prio in this case. | |
127 | */ | |
128 | if (rt_prio(current->prio)) | |
129 | ctx->prio = current->prio; | |
130 | else | |
131 | ctx->prio = current->static_prio; | |
132 | ctx->policy = current->policy; | |
ea1ae594 CH |
133 | |
134 | /* | |
91569531 LB |
135 | * TO DO: the context may be loaded, so we may need to activate |
136 | * it again on a different node. But it shouldn't hurt anything | |
137 | * to update its parameters, because we know that the scheduler | |
138 | * is not actively looking at this field, since it is not on the | |
139 | * runqueue. The context will be rescheduled on the proper node | |
140 | * if it is timesliced or preempted. | |
ea1ae594 | 141 | */ |
ea1ae594 | 142 | ctx->cpus_allowed = current->cpus_allowed; |
2cf2b3b4 CH |
143 | } |
144 | ||
145 | void spu_update_sched_info(struct spu_context *ctx) | |
146 | { | |
91569531 | 147 | int node; |
2cf2b3b4 | 148 | |
91569531 LB |
149 | if (ctx->state == SPU_STATE_RUNNABLE) { |
150 | node = ctx->spu->node; | |
e65c2f6f LB |
151 | |
152 | /* | |
153 | * Take list_mutex to sync with find_victim(). | |
154 | */ | |
91569531 LB |
155 | mutex_lock(&cbe_spu_info[node].list_mutex); |
156 | __spu_update_sched_info(ctx); | |
157 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
158 | } else { | |
159 | __spu_update_sched_info(ctx); | |
160 | } | |
2cf2b3b4 CH |
161 | } |
162 | ||
ea1ae594 | 163 | static int __node_allowed(struct spu_context *ctx, int node) |
8b3d6663 | 164 | { |
ea1ae594 CH |
165 | if (nr_cpus_node(node)) { |
166 | cpumask_t mask = node_to_cpumask(node); | |
8b3d6663 | 167 | |
ea1ae594 CH |
168 | if (cpus_intersects(mask, ctx->cpus_allowed)) |
169 | return 1; | |
170 | } | |
171 | ||
172 | return 0; | |
173 | } | |
174 | ||
175 | static int node_allowed(struct spu_context *ctx, int node) | |
176 | { | |
177 | int rval; | |
178 | ||
179 | spin_lock(&spu_prio->runq_lock); | |
180 | rval = __node_allowed(ctx, node); | |
181 | spin_unlock(&spu_prio->runq_lock); | |
182 | ||
183 | return rval; | |
8b3d6663 AB |
184 | } |
185 | ||
aed3a8c9 | 186 | void do_notify_spus_active(void) |
36aaccc1 BN |
187 | { |
188 | int node; | |
189 | ||
190 | /* | |
191 | * Wake up the active spu_contexts. | |
192 | * | |
193 | * When the awakened processes see their "notify_active" flag is set, | |
9b1d21f8 | 194 | * they will call spu_switch_notify(). |
36aaccc1 BN |
195 | */ |
196 | for_each_online_node(node) { | |
197 | struct spu *spu; | |
486acd48 CH |
198 | |
199 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
200 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { | |
201 | if (spu->alloc_state != SPU_FREE) { | |
202 | struct spu_context *ctx = spu->ctx; | |
203 | set_bit(SPU_SCHED_NOTIFY_ACTIVE, | |
204 | &ctx->sched_flags); | |
205 | mb(); | |
206 | wake_up_all(&ctx->stop_wq); | |
207 | } | |
36aaccc1 | 208 | } |
486acd48 | 209 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
36aaccc1 BN |
210 | } |
211 | } | |
212 | ||
202557d2 CH |
213 | /** |
214 | * spu_bind_context - bind spu context to physical spu | |
215 | * @spu: physical spu to bind to | |
216 | * @ctx: context to bind | |
217 | */ | |
218 | static void spu_bind_context(struct spu *spu, struct spu_context *ctx) | |
8b3d6663 | 219 | { |
038200cf CH |
220 | spu_context_trace(spu_bind_context__enter, ctx, spu); |
221 | ||
27ec41d3 | 222 | spuctx_switch_state(ctx, SPU_UTIL_SYSTEM); |
e9f8a0b6 | 223 | |
aa6d5b20 AB |
224 | if (ctx->flags & SPU_CREATE_NOSCHED) |
225 | atomic_inc(&cbe_spu_info[spu->node].reserved_spus); | |
226 | ||
e9f8a0b6 CH |
227 | ctx->stats.slb_flt_base = spu->stats.slb_flt; |
228 | ctx->stats.class2_intr_base = spu->stats.class2_intr; | |
229 | ||
8b3d6663 AB |
230 | spu->ctx = ctx; |
231 | spu->flags = 0; | |
232 | ctx->spu = spu; | |
233 | ctx->ops = &spu_hw_ops; | |
234 | spu->pid = current->pid; | |
1474855d | 235 | spu->tgid = current->tgid; |
94b2a439 | 236 | spu_associate_mm(spu, ctx->owner); |
8b3d6663 AB |
237 | spu->ibox_callback = spufs_ibox_callback; |
238 | spu->wbox_callback = spufs_wbox_callback; | |
5110459f | 239 | spu->stop_callback = spufs_stop_callback; |
a33a7d73 | 240 | spu->mfc_callback = spufs_mfc_callback; |
8b3d6663 | 241 | mb(); |
5110459f | 242 | spu_unmap_mappings(ctx); |
8b3d6663 | 243 | spu_restore(&ctx->csa, spu); |
2a911f0b | 244 | spu->timestamp = jiffies; |
a68cf983 | 245 | spu_cpu_affinity_set(spu, raw_smp_processor_id()); |
86767277 | 246 | spu_switch_notify(spu, ctx); |
81998baf | 247 | ctx->state = SPU_STATE_RUNNABLE; |
27ec41d3 AD |
248 | |
249 | spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED); | |
8b3d6663 AB |
250 | } |
251 | ||
c5fc8d2a | 252 | /* |
486acd48 | 253 | * Must be used with the list_mutex held. |
c5fc8d2a AB |
254 | */ |
255 | static inline int sched_spu(struct spu *spu) | |
256 | { | |
486acd48 CH |
257 | BUG_ON(!mutex_is_locked(&cbe_spu_info[spu->node].list_mutex)); |
258 | ||
c5fc8d2a AB |
259 | return (!spu->ctx || !(spu->ctx->flags & SPU_CREATE_NOSCHED)); |
260 | } | |
261 | ||
262 | static void aff_merge_remaining_ctxs(struct spu_gang *gang) | |
263 | { | |
264 | struct spu_context *ctx; | |
265 | ||
266 | list_for_each_entry(ctx, &gang->aff_list_head, aff_list) { | |
267 | if (list_empty(&ctx->aff_list)) | |
268 | list_add(&ctx->aff_list, &gang->aff_list_head); | |
269 | } | |
270 | gang->aff_flags |= AFF_MERGED; | |
271 | } | |
272 | ||
273 | static void aff_set_offsets(struct spu_gang *gang) | |
274 | { | |
275 | struct spu_context *ctx; | |
276 | int offset; | |
277 | ||
278 | offset = -1; | |
279 | list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list, | |
280 | aff_list) { | |
281 | if (&ctx->aff_list == &gang->aff_list_head) | |
282 | break; | |
283 | ctx->aff_offset = offset--; | |
284 | } | |
285 | ||
286 | offset = 0; | |
287 | list_for_each_entry(ctx, gang->aff_ref_ctx->aff_list.prev, aff_list) { | |
288 | if (&ctx->aff_list == &gang->aff_list_head) | |
289 | break; | |
290 | ctx->aff_offset = offset++; | |
291 | } | |
292 | ||
293 | gang->aff_flags |= AFF_OFFSETS_SET; | |
294 | } | |
295 | ||
296 | static struct spu *aff_ref_location(struct spu_context *ctx, int mem_aff, | |
297 | int group_size, int lowest_offset) | |
298 | { | |
299 | struct spu *spu; | |
300 | int node, n; | |
301 | ||
302 | /* | |
303 | * TODO: A better algorithm could be used to find a good spu to be | |
304 | * used as reference location for the ctxs chain. | |
305 | */ | |
306 | node = cpu_to_node(raw_smp_processor_id()); | |
307 | for (n = 0; n < MAX_NUMNODES; n++, node++) { | |
308 | node = (node < MAX_NUMNODES) ? node : 0; | |
309 | if (!node_allowed(ctx, node)) | |
310 | continue; | |
486acd48 | 311 | mutex_lock(&cbe_spu_info[node].list_mutex); |
c5fc8d2a AB |
312 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { |
313 | if ((!mem_aff || spu->has_mem_affinity) && | |
486acd48 CH |
314 | sched_spu(spu)) { |
315 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
c5fc8d2a | 316 | return spu; |
486acd48 | 317 | } |
c5fc8d2a | 318 | } |
486acd48 | 319 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
c5fc8d2a AB |
320 | } |
321 | return NULL; | |
322 | } | |
323 | ||
324 | static void aff_set_ref_point_location(struct spu_gang *gang) | |
325 | { | |
326 | int mem_aff, gs, lowest_offset; | |
327 | struct spu_context *ctx; | |
328 | struct spu *tmp; | |
329 | ||
330 | mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM; | |
331 | lowest_offset = 0; | |
332 | gs = 0; | |
333 | ||
334 | list_for_each_entry(tmp, &gang->aff_list_head, aff_list) | |
335 | gs++; | |
336 | ||
337 | list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list, | |
338 | aff_list) { | |
339 | if (&ctx->aff_list == &gang->aff_list_head) | |
340 | break; | |
341 | lowest_offset = ctx->aff_offset; | |
342 | } | |
343 | ||
683e3ab2 AD |
344 | gang->aff_ref_spu = aff_ref_location(gang->aff_ref_ctx, mem_aff, gs, |
345 | lowest_offset); | |
c5fc8d2a AB |
346 | } |
347 | ||
486acd48 | 348 | static struct spu *ctx_location(struct spu *ref, int offset, int node) |
c5fc8d2a AB |
349 | { |
350 | struct spu *spu; | |
351 | ||
352 | spu = NULL; | |
353 | if (offset >= 0) { | |
354 | list_for_each_entry(spu, ref->aff_list.prev, aff_list) { | |
486acd48 | 355 | BUG_ON(spu->node != node); |
c5fc8d2a AB |
356 | if (offset == 0) |
357 | break; | |
358 | if (sched_spu(spu)) | |
359 | offset--; | |
360 | } | |
361 | } else { | |
362 | list_for_each_entry_reverse(spu, ref->aff_list.next, aff_list) { | |
486acd48 | 363 | BUG_ON(spu->node != node); |
c5fc8d2a AB |
364 | if (offset == 0) |
365 | break; | |
366 | if (sched_spu(spu)) | |
367 | offset++; | |
368 | } | |
369 | } | |
486acd48 | 370 | |
c5fc8d2a AB |
371 | return spu; |
372 | } | |
373 | ||
374 | /* | |
375 | * affinity_check is called each time a context is going to be scheduled. | |
376 | * It returns the spu ptr on which the context must run. | |
377 | */ | |
486acd48 | 378 | static int has_affinity(struct spu_context *ctx) |
c5fc8d2a | 379 | { |
486acd48 | 380 | struct spu_gang *gang = ctx->gang; |
c5fc8d2a AB |
381 | |
382 | if (list_empty(&ctx->aff_list)) | |
486acd48 CH |
383 | return 0; |
384 | ||
c5fc8d2a AB |
385 | if (!gang->aff_ref_spu) { |
386 | if (!(gang->aff_flags & AFF_MERGED)) | |
387 | aff_merge_remaining_ctxs(gang); | |
388 | if (!(gang->aff_flags & AFF_OFFSETS_SET)) | |
389 | aff_set_offsets(gang); | |
390 | aff_set_ref_point_location(gang); | |
391 | } | |
486acd48 CH |
392 | |
393 | return gang->aff_ref_spu != NULL; | |
c5fc8d2a AB |
394 | } |
395 | ||
202557d2 CH |
396 | /** |
397 | * spu_unbind_context - unbind spu context from physical spu | |
398 | * @spu: physical spu to unbind from | |
399 | * @ctx: context to unbind | |
202557d2 | 400 | */ |
678b2ff1 | 401 | static void spu_unbind_context(struct spu *spu, struct spu_context *ctx) |
8b3d6663 | 402 | { |
038200cf CH |
403 | spu_context_trace(spu_unbind_context__enter, ctx, spu); |
404 | ||
27ec41d3 | 405 | spuctx_switch_state(ctx, SPU_UTIL_SYSTEM); |
fe2f896d | 406 | |
aa6d5b20 AB |
407 | if (spu->ctx->flags & SPU_CREATE_NOSCHED) |
408 | atomic_dec(&cbe_spu_info[spu->node].reserved_spus); | |
36ddbb13 AD |
409 | |
410 | if (ctx->gang){ | |
411 | mutex_lock(&ctx->gang->aff_mutex); | |
412 | if (has_affinity(ctx)) { | |
413 | if (atomic_dec_and_test(&ctx->gang->aff_sched_count)) | |
414 | ctx->gang->aff_ref_spu = NULL; | |
415 | } | |
416 | mutex_unlock(&ctx->gang->aff_mutex); | |
417 | } | |
418 | ||
86767277 | 419 | spu_switch_notify(spu, NULL); |
5110459f | 420 | spu_unmap_mappings(ctx); |
8b3d6663 | 421 | spu_save(&ctx->csa, spu); |
2a911f0b | 422 | spu->timestamp = jiffies; |
8b3d6663 AB |
423 | ctx->state = SPU_STATE_SAVED; |
424 | spu->ibox_callback = NULL; | |
425 | spu->wbox_callback = NULL; | |
5110459f | 426 | spu->stop_callback = NULL; |
a33a7d73 | 427 | spu->mfc_callback = NULL; |
94b2a439 | 428 | spu_associate_mm(spu, NULL); |
8b3d6663 | 429 | spu->pid = 0; |
1474855d | 430 | spu->tgid = 0; |
8b3d6663 | 431 | ctx->ops = &spu_backing_ops; |
2a911f0b | 432 | spu->flags = 0; |
8b3d6663 | 433 | spu->ctx = NULL; |
e9f8a0b6 CH |
434 | |
435 | ctx->stats.slb_flt += | |
436 | (spu->stats.slb_flt - ctx->stats.slb_flt_base); | |
437 | ctx->stats.class2_intr += | |
438 | (spu->stats.class2_intr - ctx->stats.class2_intr_base); | |
27ec41d3 AD |
439 | |
440 | /* This maps the underlying spu state to idle */ | |
441 | spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED); | |
442 | ctx->spu = NULL; | |
8b3d6663 AB |
443 | } |
444 | ||
079cdb61 CH |
445 | /** |
446 | * spu_add_to_rq - add a context to the runqueue | |
447 | * @ctx: context to add | |
448 | */ | |
4e0f4ed0 | 449 | static void __spu_add_to_rq(struct spu_context *ctx) |
8b3d6663 | 450 | { |
27449971 CH |
451 | /* |
452 | * Unfortunately this code path can be called from multiple threads | |
453 | * on behalf of a single context due to the way the problem state | |
454 | * mmap support works. | |
455 | * | |
456 | * Fortunately we need to wake up all these threads at the same time | |
457 | * and can simply skip the runqueue addition for every but the first | |
458 | * thread getting into this codepath. | |
459 | * | |
460 | * It's still quite hacky, and long-term we should proxy all other | |
461 | * threads through the owner thread so that spu_run is in control | |
462 | * of all the scheduling activity for a given context. | |
463 | */ | |
464 | if (list_empty(&ctx->rq)) { | |
465 | list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]); | |
466 | set_bit(ctx->prio, spu_prio->bitmap); | |
467 | if (!spu_prio->nr_waiting++) | |
468 | __mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK); | |
469 | } | |
2a911f0b | 470 | } |
5110459f | 471 | |
e65c2f6f LB |
472 | static void spu_add_to_rq(struct spu_context *ctx) |
473 | { | |
474 | spin_lock(&spu_prio->runq_lock); | |
475 | __spu_add_to_rq(ctx); | |
476 | spin_unlock(&spu_prio->runq_lock); | |
477 | } | |
478 | ||
4e0f4ed0 | 479 | static void __spu_del_from_rq(struct spu_context *ctx) |
a475c2f4 | 480 | { |
4e0f4ed0 LB |
481 | int prio = ctx->prio; |
482 | ||
65de66f0 | 483 | if (!list_empty(&ctx->rq)) { |
c77239b8 CH |
484 | if (!--spu_prio->nr_waiting) |
485 | del_timer(&spusched_timer); | |
a475c2f4 | 486 | list_del_init(&ctx->rq); |
c77239b8 CH |
487 | |
488 | if (list_empty(&spu_prio->runq[prio])) | |
489 | clear_bit(prio, spu_prio->bitmap); | |
65de66f0 | 490 | } |
079cdb61 | 491 | } |
a68cf983 | 492 | |
e65c2f6f LB |
493 | void spu_del_from_rq(struct spu_context *ctx) |
494 | { | |
495 | spin_lock(&spu_prio->runq_lock); | |
496 | __spu_del_from_rq(ctx); | |
497 | spin_unlock(&spu_prio->runq_lock); | |
498 | } | |
499 | ||
079cdb61 | 500 | static void spu_prio_wait(struct spu_context *ctx) |
8b3d6663 | 501 | { |
a68cf983 | 502 | DEFINE_WAIT(wait); |
8b3d6663 | 503 | |
e65c2f6f LB |
504 | /* |
505 | * The caller must explicitly wait for a context to be loaded | |
506 | * if the nosched flag is set. If NOSCHED is not set, the caller | |
507 | * queues the context and waits for an spu event or error. | |
508 | */ | |
509 | BUG_ON(!(ctx->flags & SPU_CREATE_NOSCHED)); | |
510 | ||
4e0f4ed0 | 511 | spin_lock(&spu_prio->runq_lock); |
079cdb61 | 512 | prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE); |
a68cf983 | 513 | if (!signal_pending(current)) { |
4e0f4ed0 LB |
514 | __spu_add_to_rq(ctx); |
515 | spin_unlock(&spu_prio->runq_lock); | |
650f8b02 | 516 | mutex_unlock(&ctx->state_mutex); |
a68cf983 | 517 | schedule(); |
650f8b02 | 518 | mutex_lock(&ctx->state_mutex); |
4e0f4ed0 LB |
519 | spin_lock(&spu_prio->runq_lock); |
520 | __spu_del_from_rq(ctx); | |
8b3d6663 | 521 | } |
4e0f4ed0 | 522 | spin_unlock(&spu_prio->runq_lock); |
079cdb61 CH |
523 | __set_current_state(TASK_RUNNING); |
524 | remove_wait_queue(&ctx->stop_wq, &wait); | |
8b3d6663 AB |
525 | } |
526 | ||
079cdb61 | 527 | static struct spu *spu_get_idle(struct spu_context *ctx) |
a68cf983 | 528 | { |
36ddbb13 | 529 | struct spu *spu, *aff_ref_spu; |
486acd48 CH |
530 | int node, n; |
531 | ||
038200cf CH |
532 | spu_context_nospu_trace(spu_get_idle__enter, ctx); |
533 | ||
36ddbb13 AD |
534 | if (ctx->gang) { |
535 | mutex_lock(&ctx->gang->aff_mutex); | |
536 | if (has_affinity(ctx)) { | |
537 | aff_ref_spu = ctx->gang->aff_ref_spu; | |
538 | atomic_inc(&ctx->gang->aff_sched_count); | |
539 | mutex_unlock(&ctx->gang->aff_mutex); | |
540 | node = aff_ref_spu->node; | |
541 | ||
542 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
543 | spu = ctx_location(aff_ref_spu, ctx->aff_offset, node); | |
544 | if (spu && spu->alloc_state == SPU_FREE) | |
545 | goto found; | |
546 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
a68cf983 | 547 | |
36ddbb13 AD |
548 | mutex_lock(&ctx->gang->aff_mutex); |
549 | if (atomic_dec_and_test(&ctx->gang->aff_sched_count)) | |
550 | ctx->gang->aff_ref_spu = NULL; | |
551 | mutex_unlock(&ctx->gang->aff_mutex); | |
038200cf | 552 | goto not_found; |
36ddbb13 AD |
553 | } |
554 | mutex_unlock(&ctx->gang->aff_mutex); | |
555 | } | |
486acd48 | 556 | node = cpu_to_node(raw_smp_processor_id()); |
a68cf983 MN |
557 | for (n = 0; n < MAX_NUMNODES; n++, node++) { |
558 | node = (node < MAX_NUMNODES) ? node : 0; | |
ea1ae594 | 559 | if (!node_allowed(ctx, node)) |
a68cf983 | 560 | continue; |
486acd48 CH |
561 | |
562 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
563 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { | |
564 | if (spu->alloc_state == SPU_FREE) | |
565 | goto found; | |
566 | } | |
567 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
a68cf983 | 568 | } |
486acd48 | 569 | |
038200cf CH |
570 | not_found: |
571 | spu_context_nospu_trace(spu_get_idle__not_found, ctx); | |
486acd48 CH |
572 | return NULL; |
573 | ||
574 | found: | |
575 | spu->alloc_state = SPU_USED; | |
576 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
038200cf | 577 | spu_context_trace(spu_get_idle__found, ctx, spu); |
486acd48 | 578 | spu_init_channels(spu); |
a68cf983 MN |
579 | return spu; |
580 | } | |
8b3d6663 | 581 | |
52f04fcf CH |
582 | /** |
583 | * find_victim - find a lower priority context to preempt | |
584 | * @ctx: canidate context for running | |
585 | * | |
586 | * Returns the freed physical spu to run the new context on. | |
587 | */ | |
588 | static struct spu *find_victim(struct spu_context *ctx) | |
589 | { | |
590 | struct spu_context *victim = NULL; | |
591 | struct spu *spu; | |
592 | int node, n; | |
593 | ||
038200cf CH |
594 | spu_context_nospu_trace(spu_find_vitim__enter, ctx); |
595 | ||
52f04fcf CH |
596 | /* |
597 | * Look for a possible preemption candidate on the local node first. | |
598 | * If there is no candidate look at the other nodes. This isn't | |
9b1d21f8 | 599 | * exactly fair, but so far the whole spu scheduler tries to keep |
52f04fcf CH |
600 | * a strong node affinity. We might want to fine-tune this in |
601 | * the future. | |
602 | */ | |
603 | restart: | |
604 | node = cpu_to_node(raw_smp_processor_id()); | |
605 | for (n = 0; n < MAX_NUMNODES; n++, node++) { | |
606 | node = (node < MAX_NUMNODES) ? node : 0; | |
ea1ae594 | 607 | if (!node_allowed(ctx, node)) |
52f04fcf CH |
608 | continue; |
609 | ||
486acd48 CH |
610 | mutex_lock(&cbe_spu_info[node].list_mutex); |
611 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { | |
52f04fcf CH |
612 | struct spu_context *tmp = spu->ctx; |
613 | ||
c0e7b4aa | 614 | if (tmp && tmp->prio > ctx->prio && |
e65c2f6f | 615 | !(tmp->flags & SPU_CREATE_NOSCHED) && |
fe443ef2 | 616 | (!victim || tmp->prio > victim->prio)) |
52f04fcf CH |
617 | victim = spu->ctx; |
618 | } | |
486acd48 | 619 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
52f04fcf CH |
620 | |
621 | if (victim) { | |
622 | /* | |
623 | * This nests ctx->state_mutex, but we always lock | |
624 | * higher priority contexts before lower priority | |
625 | * ones, so this is safe until we introduce | |
626 | * priority inheritance schemes. | |
91569531 LB |
627 | * |
628 | * XXX if the highest priority context is locked, | |
629 | * this can loop a long time. Might be better to | |
630 | * look at another context or give up after X retries. | |
52f04fcf CH |
631 | */ |
632 | if (!mutex_trylock(&victim->state_mutex)) { | |
633 | victim = NULL; | |
634 | goto restart; | |
635 | } | |
636 | ||
637 | spu = victim->spu; | |
b192541b | 638 | if (!spu || victim->prio <= ctx->prio) { |
52f04fcf CH |
639 | /* |
640 | * This race can happen because we've dropped | |
b192541b | 641 | * the active list mutex. Not a problem, just |
52f04fcf CH |
642 | * restart the search. |
643 | */ | |
644 | mutex_unlock(&victim->state_mutex); | |
645 | victim = NULL; | |
646 | goto restart; | |
647 | } | |
486acd48 | 648 | |
038200cf CH |
649 | spu_context_trace(__spu_deactivate__unload, ctx, spu); |
650 | ||
486acd48 CH |
651 | mutex_lock(&cbe_spu_info[node].list_mutex); |
652 | cbe_spu_info[node].nr_active--; | |
c0e7b4aa | 653 | spu_unbind_context(spu, victim); |
486acd48 CH |
654 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
655 | ||
e9f8a0b6 | 656 | victim->stats.invol_ctx_switch++; |
fe2f896d | 657 | spu->stats.invol_ctx_switch++; |
e65c2f6f LB |
658 | spu_add_to_rq(victim); |
659 | ||
52f04fcf | 660 | mutex_unlock(&victim->state_mutex); |
e65c2f6f | 661 | |
52f04fcf CH |
662 | return spu; |
663 | } | |
664 | } | |
665 | ||
666 | return NULL; | |
667 | } | |
668 | ||
e65c2f6f LB |
669 | static void __spu_schedule(struct spu *spu, struct spu_context *ctx) |
670 | { | |
671 | int node = spu->node; | |
672 | int success = 0; | |
673 | ||
674 | spu_set_timeslice(ctx); | |
675 | ||
676 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
677 | if (spu->ctx == NULL) { | |
678 | spu_bind_context(spu, ctx); | |
679 | cbe_spu_info[node].nr_active++; | |
680 | spu->alloc_state = SPU_USED; | |
681 | success = 1; | |
682 | } | |
683 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
684 | ||
685 | if (success) | |
686 | wake_up_all(&ctx->run_wq); | |
687 | else | |
688 | spu_add_to_rq(ctx); | |
689 | } | |
690 | ||
691 | static void spu_schedule(struct spu *spu, struct spu_context *ctx) | |
692 | { | |
c9101bdb CH |
693 | /* not a candidate for interruptible because it's called either |
694 | from the scheduler thread or from spu_deactivate */ | |
695 | mutex_lock(&ctx->state_mutex); | |
e65c2f6f LB |
696 | __spu_schedule(spu, ctx); |
697 | spu_release(ctx); | |
698 | } | |
699 | ||
700 | static void spu_unschedule(struct spu *spu, struct spu_context *ctx) | |
701 | { | |
702 | int node = spu->node; | |
703 | ||
704 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
705 | cbe_spu_info[node].nr_active--; | |
706 | spu->alloc_state = SPU_FREE; | |
707 | spu_unbind_context(spu, ctx); | |
708 | ctx->stats.invol_ctx_switch++; | |
709 | spu->stats.invol_ctx_switch++; | |
710 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
711 | } | |
712 | ||
079cdb61 CH |
713 | /** |
714 | * spu_activate - find a free spu for a context and execute it | |
715 | * @ctx: spu context to schedule | |
716 | * @flags: flags (currently ignored) | |
717 | * | |
08873095 | 718 | * Tries to find a free spu to run @ctx. If no free spu is available |
079cdb61 CH |
719 | * add the context to the runqueue so it gets woken up once an spu |
720 | * is available. | |
721 | */ | |
26bec673 | 722 | int spu_activate(struct spu_context *ctx, unsigned long flags) |
8b3d6663 | 723 | { |
e65c2f6f | 724 | struct spu *spu; |
079cdb61 | 725 | |
e65c2f6f LB |
726 | /* |
727 | * If there are multiple threads waiting for a single context | |
728 | * only one actually binds the context while the others will | |
729 | * only be able to acquire the state_mutex once the context | |
730 | * already is in runnable state. | |
731 | */ | |
732 | if (ctx->spu) | |
733 | return 0; | |
27449971 | 734 | |
e65c2f6f LB |
735 | spu_activate_top: |
736 | if (signal_pending(current)) | |
737 | return -ERESTARTSYS; | |
486acd48 | 738 | |
e65c2f6f LB |
739 | spu = spu_get_idle(ctx); |
740 | /* | |
741 | * If this is a realtime thread we try to get it running by | |
742 | * preempting a lower priority thread. | |
743 | */ | |
744 | if (!spu && rt_prio(ctx->prio)) | |
745 | spu = find_victim(ctx); | |
746 | if (spu) { | |
747 | unsigned long runcntl; | |
748 | ||
749 | runcntl = ctx->ops->runcntl_read(ctx); | |
750 | __spu_schedule(spu, ctx); | |
751 | if (runcntl & SPU_RUNCNTL_RUNNABLE) | |
752 | spuctx_switch_state(ctx, SPU_UTIL_USER); | |
079cdb61 | 753 | |
e65c2f6f LB |
754 | return 0; |
755 | } | |
756 | ||
757 | if (ctx->flags & SPU_CREATE_NOSCHED) { | |
50b520d4 | 758 | spu_prio_wait(ctx); |
e65c2f6f LB |
759 | goto spu_activate_top; |
760 | } | |
761 | ||
762 | spu_add_to_rq(ctx); | |
079cdb61 | 763 | |
e65c2f6f | 764 | return 0; |
8b3d6663 AB |
765 | } |
766 | ||
bb5db29a CH |
767 | /** |
768 | * grab_runnable_context - try to find a runnable context | |
769 | * | |
770 | * Remove the highest priority context on the runqueue and return it | |
771 | * to the caller. Returns %NULL if no runnable context was found. | |
772 | */ | |
ea1ae594 | 773 | static struct spu_context *grab_runnable_context(int prio, int node) |
bb5db29a | 774 | { |
ea1ae594 | 775 | struct spu_context *ctx; |
bb5db29a CH |
776 | int best; |
777 | ||
778 | spin_lock(&spu_prio->runq_lock); | |
7e90b749 | 779 | best = find_first_bit(spu_prio->bitmap, prio); |
ea1ae594 | 780 | while (best < prio) { |
bb5db29a CH |
781 | struct list_head *rq = &spu_prio->runq[best]; |
782 | ||
ea1ae594 CH |
783 | list_for_each_entry(ctx, rq, rq) { |
784 | /* XXX(hch): check for affinity here aswell */ | |
785 | if (__node_allowed(ctx, node)) { | |
786 | __spu_del_from_rq(ctx); | |
787 | goto found; | |
788 | } | |
789 | } | |
790 | best++; | |
bb5db29a | 791 | } |
ea1ae594 CH |
792 | ctx = NULL; |
793 | found: | |
bb5db29a | 794 | spin_unlock(&spu_prio->runq_lock); |
bb5db29a CH |
795 | return ctx; |
796 | } | |
797 | ||
798 | static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio) | |
799 | { | |
800 | struct spu *spu = ctx->spu; | |
801 | struct spu_context *new = NULL; | |
802 | ||
803 | if (spu) { | |
ea1ae594 | 804 | new = grab_runnable_context(max_prio, spu->node); |
bb5db29a | 805 | if (new || force) { |
e65c2f6f LB |
806 | spu_unschedule(spu, ctx); |
807 | if (new) { | |
808 | if (new->flags & SPU_CREATE_NOSCHED) | |
809 | wake_up(&new->stop_wq); | |
810 | else { | |
811 | spu_release(ctx); | |
812 | spu_schedule(spu, new); | |
c9101bdb CH |
813 | /* this one can't easily be made |
814 | interruptible */ | |
815 | mutex_lock(&ctx->state_mutex); | |
e65c2f6f LB |
816 | } |
817 | } | |
bb5db29a | 818 | } |
bb5db29a CH |
819 | } |
820 | ||
821 | return new != NULL; | |
822 | } | |
823 | ||
678b2ff1 CH |
824 | /** |
825 | * spu_deactivate - unbind a context from it's physical spu | |
826 | * @ctx: spu context to unbind | |
827 | * | |
828 | * Unbind @ctx from the physical spu it is running on and schedule | |
829 | * the highest priority context to run on the freed physical spu. | |
830 | */ | |
8b3d6663 AB |
831 | void spu_deactivate(struct spu_context *ctx) |
832 | { | |
038200cf | 833 | spu_context_nospu_trace(spu_deactivate__enter, ctx); |
bb5db29a | 834 | __spu_deactivate(ctx, 1, MAX_PRIO); |
8b3d6663 AB |
835 | } |
836 | ||
ae7b4c52 | 837 | /** |
1474855d | 838 | * spu_yield - yield a physical spu if others are waiting |
ae7b4c52 CH |
839 | * @ctx: spu context to yield |
840 | * | |
841 | * Check if there is a higher priority context waiting and if yes | |
842 | * unbind @ctx from the physical spu and schedule the highest | |
843 | * priority context to run on the freed physical spu instead. | |
844 | */ | |
8b3d6663 AB |
845 | void spu_yield(struct spu_context *ctx) |
846 | { | |
038200cf | 847 | spu_context_nospu_trace(spu_yield__enter, ctx); |
e5c0b9ec CH |
848 | if (!(ctx->flags & SPU_CREATE_NOSCHED)) { |
849 | mutex_lock(&ctx->state_mutex); | |
27ec41d3 | 850 | __spu_deactivate(ctx, 0, MAX_PRIO); |
e5c0b9ec CH |
851 | mutex_unlock(&ctx->state_mutex); |
852 | } | |
bb5db29a | 853 | } |
8b3d6663 | 854 | |
486acd48 | 855 | static noinline void spusched_tick(struct spu_context *ctx) |
bb5db29a | 856 | { |
e65c2f6f LB |
857 | struct spu_context *new = NULL; |
858 | struct spu *spu = NULL; | |
859 | u32 status; | |
860 | ||
c9101bdb CH |
861 | if (spu_acquire(ctx)) |
862 | BUG(); /* a kernel thread never has signals pending */ | |
e65c2f6f LB |
863 | |
864 | if (ctx->state != SPU_STATE_RUNNABLE) | |
865 | goto out; | |
866 | if (spu_stopped(ctx, &status)) | |
867 | goto out; | |
df09cf3e | 868 | if (ctx->flags & SPU_CREATE_NOSCHED) |
e65c2f6f | 869 | goto out; |
df09cf3e | 870 | if (ctx->policy == SCHED_FIFO) |
e65c2f6f | 871 | goto out; |
df09cf3e CH |
872 | |
873 | if (--ctx->time_slice) | |
e65c2f6f | 874 | goto out; |
bb5db29a | 875 | |
e65c2f6f | 876 | spu = ctx->spu; |
038200cf CH |
877 | |
878 | spu_context_trace(spusched_tick__preempt, ctx, spu); | |
879 | ||
e65c2f6f LB |
880 | new = grab_runnable_context(ctx->prio + 1, spu->node); |
881 | if (new) { | |
882 | spu_unschedule(spu, ctx); | |
883 | spu_add_to_rq(ctx); | |
bb5db29a | 884 | } else { |
038200cf | 885 | spu_context_nospu_trace(spusched_tick__newslice, ctx); |
37901802 | 886 | ctx->time_slice++; |
8b3d6663 | 887 | } |
e65c2f6f LB |
888 | out: |
889 | spu_release(ctx); | |
890 | ||
891 | if (new) | |
892 | spu_schedule(spu, new); | |
8b3d6663 AB |
893 | } |
894 | ||
65de66f0 CH |
895 | /** |
896 | * count_active_contexts - count nr of active tasks | |
897 | * | |
898 | * Return the number of tasks currently running or waiting to run. | |
899 | * | |
486acd48 | 900 | * Note that we don't take runq_lock / list_mutex here. Reading |
65de66f0 CH |
901 | * a single 32bit value is atomic on powerpc, and we don't care |
902 | * about memory ordering issues here. | |
903 | */ | |
904 | static unsigned long count_active_contexts(void) | |
905 | { | |
906 | int nr_active = 0, node; | |
907 | ||
908 | for (node = 0; node < MAX_NUMNODES; node++) | |
486acd48 | 909 | nr_active += cbe_spu_info[node].nr_active; |
65de66f0 CH |
910 | nr_active += spu_prio->nr_waiting; |
911 | ||
912 | return nr_active; | |
913 | } | |
914 | ||
915 | /** | |
90608a29 | 916 | * spu_calc_load - update the avenrun load estimates. |
65de66f0 CH |
917 | * |
918 | * No locking against reading these values from userspace, as for | |
919 | * the CPU loadavg code. | |
920 | */ | |
90608a29 | 921 | static void spu_calc_load(void) |
65de66f0 CH |
922 | { |
923 | unsigned long active_tasks; /* fixed-point */ | |
90608a29 AL |
924 | |
925 | active_tasks = count_active_contexts() * FIXED_1; | |
926 | CALC_LOAD(spu_avenrun[0], EXP_1, active_tasks); | |
927 | CALC_LOAD(spu_avenrun[1], EXP_5, active_tasks); | |
928 | CALC_LOAD(spu_avenrun[2], EXP_15, active_tasks); | |
65de66f0 CH |
929 | } |
930 | ||
37901802 CH |
931 | static void spusched_wake(unsigned long data) |
932 | { | |
933 | mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK); | |
934 | wake_up_process(spusched_task); | |
90608a29 AL |
935 | } |
936 | ||
937 | static void spuloadavg_wake(unsigned long data) | |
938 | { | |
939 | mod_timer(&spuloadavg_timer, jiffies + LOAD_FREQ); | |
940 | spu_calc_load(); | |
37901802 CH |
941 | } |
942 | ||
943 | static int spusched_thread(void *unused) | |
944 | { | |
486acd48 | 945 | struct spu *spu; |
37901802 CH |
946 | int node; |
947 | ||
37901802 CH |
948 | while (!kthread_should_stop()) { |
949 | set_current_state(TASK_INTERRUPTIBLE); | |
950 | schedule(); | |
951 | for (node = 0; node < MAX_NUMNODES; node++) { | |
e65c2f6f LB |
952 | struct mutex *mtx = &cbe_spu_info[node].list_mutex; |
953 | ||
954 | mutex_lock(mtx); | |
955 | list_for_each_entry(spu, &cbe_spu_info[node].spus, | |
956 | cbe_list) { | |
957 | struct spu_context *ctx = spu->ctx; | |
958 | ||
959 | if (ctx) { | |
960 | mutex_unlock(mtx); | |
961 | spusched_tick(ctx); | |
962 | mutex_lock(mtx); | |
963 | } | |
964 | } | |
965 | mutex_unlock(mtx); | |
37901802 CH |
966 | } |
967 | } | |
968 | ||
37901802 CH |
969 | return 0; |
970 | } | |
971 | ||
7cd58e43 JK |
972 | void spuctx_switch_state(struct spu_context *ctx, |
973 | enum spu_utilization_state new_state) | |
974 | { | |
975 | unsigned long long curtime; | |
976 | signed long long delta; | |
977 | struct timespec ts; | |
978 | struct spu *spu; | |
979 | enum spu_utilization_state old_state; | |
980 | ||
981 | ktime_get_ts(&ts); | |
982 | curtime = timespec_to_ns(&ts); | |
983 | delta = curtime - ctx->stats.tstamp; | |
984 | ||
985 | WARN_ON(!mutex_is_locked(&ctx->state_mutex)); | |
986 | WARN_ON(delta < 0); | |
987 | ||
988 | spu = ctx->spu; | |
989 | old_state = ctx->stats.util_state; | |
990 | ctx->stats.util_state = new_state; | |
991 | ctx->stats.tstamp = curtime; | |
992 | ||
993 | /* | |
994 | * Update the physical SPU utilization statistics. | |
995 | */ | |
996 | if (spu) { | |
997 | ctx->stats.times[old_state] += delta; | |
998 | spu->stats.times[old_state] += delta; | |
999 | spu->stats.util_state = new_state; | |
1000 | spu->stats.tstamp = curtime; | |
1001 | } | |
1002 | } | |
1003 | ||
65de66f0 CH |
1004 | #define LOAD_INT(x) ((x) >> FSHIFT) |
1005 | #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100) | |
1006 | ||
1007 | static int show_spu_loadavg(struct seq_file *s, void *private) | |
1008 | { | |
1009 | int a, b, c; | |
1010 | ||
1011 | a = spu_avenrun[0] + (FIXED_1/200); | |
1012 | b = spu_avenrun[1] + (FIXED_1/200); | |
1013 | c = spu_avenrun[2] + (FIXED_1/200); | |
1014 | ||
1015 | /* | |
1016 | * Note that last_pid doesn't really make much sense for the | |
9b1d21f8 | 1017 | * SPU loadavg (it even seems very odd on the CPU side...), |
65de66f0 CH |
1018 | * but we include it here to have a 100% compatible interface. |
1019 | */ | |
1020 | seq_printf(s, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n", | |
1021 | LOAD_INT(a), LOAD_FRAC(a), | |
1022 | LOAD_INT(b), LOAD_FRAC(b), | |
1023 | LOAD_INT(c), LOAD_FRAC(c), | |
1024 | count_active_contexts(), | |
1025 | atomic_read(&nr_spu_contexts), | |
1026 | current->nsproxy->pid_ns->last_pid); | |
1027 | return 0; | |
1028 | } | |
1029 | ||
1030 | static int spu_loadavg_open(struct inode *inode, struct file *file) | |
1031 | { | |
1032 | return single_open(file, show_spu_loadavg, NULL); | |
1033 | } | |
1034 | ||
1035 | static const struct file_operations spu_loadavg_fops = { | |
1036 | .open = spu_loadavg_open, | |
1037 | .read = seq_read, | |
1038 | .llseek = seq_lseek, | |
1039 | .release = single_release, | |
1040 | }; | |
1041 | ||
8b3d6663 AB |
1042 | int __init spu_sched_init(void) |
1043 | { | |
65de66f0 CH |
1044 | struct proc_dir_entry *entry; |
1045 | int err = -ENOMEM, i; | |
8b3d6663 | 1046 | |
a68cf983 | 1047 | spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL); |
37901802 | 1048 | if (!spu_prio) |
65de66f0 | 1049 | goto out; |
37901802 | 1050 | |
8b3d6663 | 1051 | for (i = 0; i < MAX_PRIO; i++) { |
079cdb61 | 1052 | INIT_LIST_HEAD(&spu_prio->runq[i]); |
a68cf983 | 1053 | __clear_bit(i, spu_prio->bitmap); |
8b3d6663 | 1054 | } |
079cdb61 | 1055 | spin_lock_init(&spu_prio->runq_lock); |
37901802 | 1056 | |
c77239b8 | 1057 | setup_timer(&spusched_timer, spusched_wake, 0); |
90608a29 | 1058 | setup_timer(&spuloadavg_timer, spuloadavg_wake, 0); |
c77239b8 | 1059 | |
37901802 CH |
1060 | spusched_task = kthread_run(spusched_thread, NULL, "spusched"); |
1061 | if (IS_ERR(spusched_task)) { | |
65de66f0 CH |
1062 | err = PTR_ERR(spusched_task); |
1063 | goto out_free_spu_prio; | |
37901802 | 1064 | } |
f3f59bec | 1065 | |
90608a29 AL |
1066 | mod_timer(&spuloadavg_timer, 0); |
1067 | ||
65de66f0 CH |
1068 | entry = create_proc_entry("spu_loadavg", 0, NULL); |
1069 | if (!entry) | |
1070 | goto out_stop_kthread; | |
1071 | entry->proc_fops = &spu_loadavg_fops; | |
1072 | ||
f3f59bec JK |
1073 | pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n", |
1074 | SPUSCHED_TICK, MIN_SPU_TIMESLICE, DEF_SPU_TIMESLICE); | |
8b3d6663 | 1075 | return 0; |
37901802 | 1076 | |
65de66f0 CH |
1077 | out_stop_kthread: |
1078 | kthread_stop(spusched_task); | |
1079 | out_free_spu_prio: | |
1080 | kfree(spu_prio); | |
1081 | out: | |
1082 | return err; | |
8b3d6663 AB |
1083 | } |
1084 | ||
d1450317 | 1085 | void spu_sched_exit(void) |
8b3d6663 | 1086 | { |
486acd48 | 1087 | struct spu *spu; |
a68cf983 MN |
1088 | int node; |
1089 | ||
65de66f0 CH |
1090 | remove_proc_entry("spu_loadavg", NULL); |
1091 | ||
c77239b8 | 1092 | del_timer_sync(&spusched_timer); |
90608a29 | 1093 | del_timer_sync(&spuloadavg_timer); |
37901802 CH |
1094 | kthread_stop(spusched_task); |
1095 | ||
a68cf983 | 1096 | for (node = 0; node < MAX_NUMNODES; node++) { |
486acd48 CH |
1097 | mutex_lock(&cbe_spu_info[node].list_mutex); |
1098 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) | |
1099 | if (spu->alloc_state != SPU_FREE) | |
1100 | spu->alloc_state = SPU_FREE; | |
1101 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
8b3d6663 | 1102 | } |
a68cf983 | 1103 | kfree(spu_prio); |
8b3d6663 | 1104 | } |