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; |
7a214200 LB |
143 | |
144 | /* Save the current cpu id for spu interrupt routing. */ | |
145 | ctx->last_ran = raw_smp_processor_id(); | |
2cf2b3b4 CH |
146 | } |
147 | ||
148 | void spu_update_sched_info(struct spu_context *ctx) | |
149 | { | |
91569531 | 150 | int node; |
2cf2b3b4 | 151 | |
91569531 LB |
152 | if (ctx->state == SPU_STATE_RUNNABLE) { |
153 | node = ctx->spu->node; | |
e65c2f6f LB |
154 | |
155 | /* | |
156 | * Take list_mutex to sync with find_victim(). | |
157 | */ | |
91569531 LB |
158 | mutex_lock(&cbe_spu_info[node].list_mutex); |
159 | __spu_update_sched_info(ctx); | |
160 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
161 | } else { | |
162 | __spu_update_sched_info(ctx); | |
163 | } | |
2cf2b3b4 CH |
164 | } |
165 | ||
ea1ae594 | 166 | static int __node_allowed(struct spu_context *ctx, int node) |
8b3d6663 | 167 | { |
ea1ae594 | 168 | if (nr_cpus_node(node)) { |
86c6f274 | 169 | const struct cpumask *mask = cpumask_of_node(node); |
8b3d6663 | 170 | |
86c6f274 | 171 | if (cpumask_intersects(mask, &ctx->cpus_allowed)) |
ea1ae594 CH |
172 | return 1; |
173 | } | |
174 | ||
175 | return 0; | |
176 | } | |
177 | ||
178 | static int node_allowed(struct spu_context *ctx, int node) | |
179 | { | |
180 | int rval; | |
181 | ||
182 | spin_lock(&spu_prio->runq_lock); | |
183 | rval = __node_allowed(ctx, node); | |
184 | spin_unlock(&spu_prio->runq_lock); | |
185 | ||
186 | return rval; | |
8b3d6663 AB |
187 | } |
188 | ||
aed3a8c9 | 189 | void do_notify_spus_active(void) |
36aaccc1 BN |
190 | { |
191 | int node; | |
192 | ||
193 | /* | |
194 | * Wake up the active spu_contexts. | |
195 | * | |
196 | * When the awakened processes see their "notify_active" flag is set, | |
9b1d21f8 | 197 | * they will call spu_switch_notify(). |
36aaccc1 BN |
198 | */ |
199 | for_each_online_node(node) { | |
200 | struct spu *spu; | |
486acd48 CH |
201 | |
202 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
203 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { | |
204 | if (spu->alloc_state != SPU_FREE) { | |
205 | struct spu_context *ctx = spu->ctx; | |
206 | set_bit(SPU_SCHED_NOTIFY_ACTIVE, | |
207 | &ctx->sched_flags); | |
208 | mb(); | |
209 | wake_up_all(&ctx->stop_wq); | |
210 | } | |
36aaccc1 | 211 | } |
486acd48 | 212 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
36aaccc1 BN |
213 | } |
214 | } | |
215 | ||
202557d2 CH |
216 | /** |
217 | * spu_bind_context - bind spu context to physical spu | |
218 | * @spu: physical spu to bind to | |
219 | * @ctx: context to bind | |
220 | */ | |
221 | static void spu_bind_context(struct spu *spu, struct spu_context *ctx) | |
8b3d6663 | 222 | { |
038200cf CH |
223 | spu_context_trace(spu_bind_context__enter, ctx, spu); |
224 | ||
27ec41d3 | 225 | spuctx_switch_state(ctx, SPU_UTIL_SYSTEM); |
e9f8a0b6 | 226 | |
aa6d5b20 AB |
227 | if (ctx->flags & SPU_CREATE_NOSCHED) |
228 | atomic_inc(&cbe_spu_info[spu->node].reserved_spus); | |
229 | ||
e9f8a0b6 CH |
230 | ctx->stats.slb_flt_base = spu->stats.slb_flt; |
231 | ctx->stats.class2_intr_base = spu->stats.class2_intr; | |
232 | ||
2c911a14 LB |
233 | spu_associate_mm(spu, ctx->owner); |
234 | ||
235 | spin_lock_irq(&spu->register_lock); | |
8b3d6663 AB |
236 | spu->ctx = ctx; |
237 | spu->flags = 0; | |
238 | ctx->spu = spu; | |
239 | ctx->ops = &spu_hw_ops; | |
240 | spu->pid = current->pid; | |
1474855d | 241 | spu->tgid = current->tgid; |
8b3d6663 AB |
242 | spu->ibox_callback = spufs_ibox_callback; |
243 | spu->wbox_callback = spufs_wbox_callback; | |
5110459f | 244 | spu->stop_callback = spufs_stop_callback; |
a33a7d73 | 245 | spu->mfc_callback = spufs_mfc_callback; |
2c911a14 LB |
246 | spin_unlock_irq(&spu->register_lock); |
247 | ||
5110459f | 248 | spu_unmap_mappings(ctx); |
2c911a14 | 249 | |
5158e9b5 | 250 | spu_switch_log_notify(spu, ctx, SWITCH_LOG_START, 0); |
8b3d6663 | 251 | spu_restore(&ctx->csa, spu); |
2a911f0b | 252 | spu->timestamp = jiffies; |
86767277 | 253 | spu_switch_notify(spu, ctx); |
81998baf | 254 | ctx->state = SPU_STATE_RUNNABLE; |
27ec41d3 | 255 | |
2a58aa33 | 256 | spuctx_switch_state(ctx, SPU_UTIL_USER); |
8b3d6663 AB |
257 | } |
258 | ||
c5fc8d2a | 259 | /* |
486acd48 | 260 | * Must be used with the list_mutex held. |
c5fc8d2a AB |
261 | */ |
262 | static inline int sched_spu(struct spu *spu) | |
263 | { | |
486acd48 CH |
264 | BUG_ON(!mutex_is_locked(&cbe_spu_info[spu->node].list_mutex)); |
265 | ||
c5fc8d2a AB |
266 | return (!spu->ctx || !(spu->ctx->flags & SPU_CREATE_NOSCHED)); |
267 | } | |
268 | ||
269 | static void aff_merge_remaining_ctxs(struct spu_gang *gang) | |
270 | { | |
271 | struct spu_context *ctx; | |
272 | ||
273 | list_for_each_entry(ctx, &gang->aff_list_head, aff_list) { | |
274 | if (list_empty(&ctx->aff_list)) | |
275 | list_add(&ctx->aff_list, &gang->aff_list_head); | |
276 | } | |
277 | gang->aff_flags |= AFF_MERGED; | |
278 | } | |
279 | ||
280 | static void aff_set_offsets(struct spu_gang *gang) | |
281 | { | |
282 | struct spu_context *ctx; | |
283 | int offset; | |
284 | ||
285 | offset = -1; | |
286 | list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list, | |
287 | aff_list) { | |
288 | if (&ctx->aff_list == &gang->aff_list_head) | |
289 | break; | |
290 | ctx->aff_offset = offset--; | |
291 | } | |
292 | ||
293 | offset = 0; | |
294 | list_for_each_entry(ctx, gang->aff_ref_ctx->aff_list.prev, aff_list) { | |
295 | if (&ctx->aff_list == &gang->aff_list_head) | |
296 | break; | |
297 | ctx->aff_offset = offset++; | |
298 | } | |
299 | ||
300 | gang->aff_flags |= AFF_OFFSETS_SET; | |
301 | } | |
302 | ||
303 | static struct spu *aff_ref_location(struct spu_context *ctx, int mem_aff, | |
304 | int group_size, int lowest_offset) | |
305 | { | |
306 | struct spu *spu; | |
307 | int node, n; | |
308 | ||
309 | /* | |
310 | * TODO: A better algorithm could be used to find a good spu to be | |
311 | * used as reference location for the ctxs chain. | |
312 | */ | |
313 | node = cpu_to_node(raw_smp_processor_id()); | |
314 | for (n = 0; n < MAX_NUMNODES; n++, node++) { | |
10baa26c AD |
315 | /* |
316 | * "available_spus" counts how many spus are not potentially | |
317 | * going to be used by other affinity gangs whose reference | |
318 | * context is already in place. Although this code seeks to | |
319 | * avoid having affinity gangs with a summed amount of | |
320 | * contexts bigger than the amount of spus in the node, | |
321 | * this may happen sporadically. In this case, available_spus | |
322 | * becomes negative, which is harmless. | |
323 | */ | |
ad1ede12 AD |
324 | int available_spus; |
325 | ||
c5fc8d2a AB |
326 | node = (node < MAX_NUMNODES) ? node : 0; |
327 | if (!node_allowed(ctx, node)) | |
328 | continue; | |
ad1ede12 AD |
329 | |
330 | available_spus = 0; | |
486acd48 | 331 | mutex_lock(&cbe_spu_info[node].list_mutex); |
ad1ede12 | 332 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { |
10baa26c AD |
333 | if (spu->ctx && spu->ctx->gang && !spu->ctx->aff_offset |
334 | && spu->ctx->gang->aff_ref_spu) | |
335 | available_spus -= spu->ctx->gang->contexts; | |
336 | available_spus++; | |
ad1ede12 AD |
337 | } |
338 | if (available_spus < ctx->gang->contexts) { | |
339 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
340 | continue; | |
341 | } | |
342 | ||
c5fc8d2a AB |
343 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { |
344 | if ((!mem_aff || spu->has_mem_affinity) && | |
486acd48 CH |
345 | sched_spu(spu)) { |
346 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
c5fc8d2a | 347 | return spu; |
486acd48 | 348 | } |
c5fc8d2a | 349 | } |
486acd48 | 350 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
c5fc8d2a AB |
351 | } |
352 | return NULL; | |
353 | } | |
354 | ||
355 | static void aff_set_ref_point_location(struct spu_gang *gang) | |
356 | { | |
357 | int mem_aff, gs, lowest_offset; | |
358 | struct spu_context *ctx; | |
359 | struct spu *tmp; | |
360 | ||
361 | mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM; | |
362 | lowest_offset = 0; | |
363 | gs = 0; | |
364 | ||
365 | list_for_each_entry(tmp, &gang->aff_list_head, aff_list) | |
366 | gs++; | |
367 | ||
368 | list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list, | |
369 | aff_list) { | |
370 | if (&ctx->aff_list == &gang->aff_list_head) | |
371 | break; | |
372 | lowest_offset = ctx->aff_offset; | |
373 | } | |
374 | ||
683e3ab2 AD |
375 | gang->aff_ref_spu = aff_ref_location(gang->aff_ref_ctx, mem_aff, gs, |
376 | lowest_offset); | |
c5fc8d2a AB |
377 | } |
378 | ||
486acd48 | 379 | static struct spu *ctx_location(struct spu *ref, int offset, int node) |
c5fc8d2a AB |
380 | { |
381 | struct spu *spu; | |
382 | ||
383 | spu = NULL; | |
384 | if (offset >= 0) { | |
385 | list_for_each_entry(spu, ref->aff_list.prev, aff_list) { | |
486acd48 | 386 | BUG_ON(spu->node != node); |
c5fc8d2a AB |
387 | if (offset == 0) |
388 | break; | |
389 | if (sched_spu(spu)) | |
390 | offset--; | |
391 | } | |
392 | } else { | |
393 | list_for_each_entry_reverse(spu, ref->aff_list.next, aff_list) { | |
486acd48 | 394 | BUG_ON(spu->node != node); |
c5fc8d2a AB |
395 | if (offset == 0) |
396 | break; | |
397 | if (sched_spu(spu)) | |
398 | offset++; | |
399 | } | |
400 | } | |
486acd48 | 401 | |
c5fc8d2a AB |
402 | return spu; |
403 | } | |
404 | ||
405 | /* | |
406 | * affinity_check is called each time a context is going to be scheduled. | |
407 | * It returns the spu ptr on which the context must run. | |
408 | */ | |
486acd48 | 409 | static int has_affinity(struct spu_context *ctx) |
c5fc8d2a | 410 | { |
486acd48 | 411 | struct spu_gang *gang = ctx->gang; |
c5fc8d2a AB |
412 | |
413 | if (list_empty(&ctx->aff_list)) | |
486acd48 CH |
414 | return 0; |
415 | ||
0855b543 AD |
416 | if (atomic_read(&ctx->gang->aff_sched_count) == 0) |
417 | ctx->gang->aff_ref_spu = NULL; | |
418 | ||
c5fc8d2a AB |
419 | if (!gang->aff_ref_spu) { |
420 | if (!(gang->aff_flags & AFF_MERGED)) | |
421 | aff_merge_remaining_ctxs(gang); | |
422 | if (!(gang->aff_flags & AFF_OFFSETS_SET)) | |
423 | aff_set_offsets(gang); | |
424 | aff_set_ref_point_location(gang); | |
425 | } | |
486acd48 CH |
426 | |
427 | return gang->aff_ref_spu != NULL; | |
c5fc8d2a AB |
428 | } |
429 | ||
202557d2 CH |
430 | /** |
431 | * spu_unbind_context - unbind spu context from physical spu | |
432 | * @spu: physical spu to unbind from | |
433 | * @ctx: context to unbind | |
202557d2 | 434 | */ |
678b2ff1 | 435 | static void spu_unbind_context(struct spu *spu, struct spu_context *ctx) |
8b3d6663 | 436 | { |
028fda0a LB |
437 | u32 status; |
438 | ||
038200cf CH |
439 | spu_context_trace(spu_unbind_context__enter, ctx, spu); |
440 | ||
27ec41d3 | 441 | spuctx_switch_state(ctx, SPU_UTIL_SYSTEM); |
fe2f896d | 442 | |
aa6d5b20 AB |
443 | if (spu->ctx->flags & SPU_CREATE_NOSCHED) |
444 | atomic_dec(&cbe_spu_info[spu->node].reserved_spus); | |
36ddbb13 | 445 | |
0855b543 | 446 | if (ctx->gang) |
34318c25 AD |
447 | /* |
448 | * If ctx->gang->aff_sched_count is positive, SPU affinity is | |
449 | * being considered in this gang. Using atomic_dec_if_positive | |
450 | * allow us to skip an explicit check for affinity in this gang | |
451 | */ | |
0855b543 | 452 | atomic_dec_if_positive(&ctx->gang->aff_sched_count); |
36ddbb13 | 453 | |
86767277 | 454 | spu_switch_notify(spu, NULL); |
5110459f | 455 | spu_unmap_mappings(ctx); |
8b3d6663 | 456 | spu_save(&ctx->csa, spu); |
5158e9b5 | 457 | spu_switch_log_notify(spu, ctx, SWITCH_LOG_STOP, 0); |
2c911a14 LB |
458 | |
459 | spin_lock_irq(&spu->register_lock); | |
2a911f0b | 460 | spu->timestamp = jiffies; |
8b3d6663 AB |
461 | ctx->state = SPU_STATE_SAVED; |
462 | spu->ibox_callback = NULL; | |
463 | spu->wbox_callback = NULL; | |
5110459f | 464 | spu->stop_callback = NULL; |
a33a7d73 | 465 | spu->mfc_callback = NULL; |
8b3d6663 | 466 | spu->pid = 0; |
1474855d | 467 | spu->tgid = 0; |
8b3d6663 | 468 | ctx->ops = &spu_backing_ops; |
2a911f0b | 469 | spu->flags = 0; |
8b3d6663 | 470 | spu->ctx = NULL; |
2c911a14 LB |
471 | spin_unlock_irq(&spu->register_lock); |
472 | ||
473 | spu_associate_mm(spu, NULL); | |
e9f8a0b6 CH |
474 | |
475 | ctx->stats.slb_flt += | |
476 | (spu->stats.slb_flt - ctx->stats.slb_flt_base); | |
477 | ctx->stats.class2_intr += | |
478 | (spu->stats.class2_intr - ctx->stats.class2_intr_base); | |
27ec41d3 AD |
479 | |
480 | /* This maps the underlying spu state to idle */ | |
481 | spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED); | |
482 | ctx->spu = NULL; | |
028fda0a LB |
483 | |
484 | if (spu_stopped(ctx, &status)) | |
485 | wake_up_all(&ctx->stop_wq); | |
8b3d6663 AB |
486 | } |
487 | ||
079cdb61 CH |
488 | /** |
489 | * spu_add_to_rq - add a context to the runqueue | |
490 | * @ctx: context to add | |
491 | */ | |
4e0f4ed0 | 492 | static void __spu_add_to_rq(struct spu_context *ctx) |
8b3d6663 | 493 | { |
27449971 CH |
494 | /* |
495 | * Unfortunately this code path can be called from multiple threads | |
496 | * on behalf of a single context due to the way the problem state | |
497 | * mmap support works. | |
498 | * | |
499 | * Fortunately we need to wake up all these threads at the same time | |
500 | * and can simply skip the runqueue addition for every but the first | |
501 | * thread getting into this codepath. | |
502 | * | |
503 | * It's still quite hacky, and long-term we should proxy all other | |
504 | * threads through the owner thread so that spu_run is in control | |
505 | * of all the scheduling activity for a given context. | |
506 | */ | |
507 | if (list_empty(&ctx->rq)) { | |
508 | list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]); | |
509 | set_bit(ctx->prio, spu_prio->bitmap); | |
510 | if (!spu_prio->nr_waiting++) | |
511 | __mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK); | |
512 | } | |
2a911f0b | 513 | } |
5110459f | 514 | |
e65c2f6f LB |
515 | static void spu_add_to_rq(struct spu_context *ctx) |
516 | { | |
517 | spin_lock(&spu_prio->runq_lock); | |
518 | __spu_add_to_rq(ctx); | |
519 | spin_unlock(&spu_prio->runq_lock); | |
520 | } | |
521 | ||
4e0f4ed0 | 522 | static void __spu_del_from_rq(struct spu_context *ctx) |
a475c2f4 | 523 | { |
4e0f4ed0 LB |
524 | int prio = ctx->prio; |
525 | ||
65de66f0 | 526 | if (!list_empty(&ctx->rq)) { |
c77239b8 CH |
527 | if (!--spu_prio->nr_waiting) |
528 | del_timer(&spusched_timer); | |
a475c2f4 | 529 | list_del_init(&ctx->rq); |
c77239b8 CH |
530 | |
531 | if (list_empty(&spu_prio->runq[prio])) | |
532 | clear_bit(prio, spu_prio->bitmap); | |
65de66f0 | 533 | } |
079cdb61 | 534 | } |
a68cf983 | 535 | |
e65c2f6f LB |
536 | void spu_del_from_rq(struct spu_context *ctx) |
537 | { | |
538 | spin_lock(&spu_prio->runq_lock); | |
539 | __spu_del_from_rq(ctx); | |
540 | spin_unlock(&spu_prio->runq_lock); | |
541 | } | |
542 | ||
079cdb61 | 543 | static void spu_prio_wait(struct spu_context *ctx) |
8b3d6663 | 544 | { |
a68cf983 | 545 | DEFINE_WAIT(wait); |
8b3d6663 | 546 | |
e65c2f6f LB |
547 | /* |
548 | * The caller must explicitly wait for a context to be loaded | |
549 | * if the nosched flag is set. If NOSCHED is not set, the caller | |
550 | * queues the context and waits for an spu event or error. | |
551 | */ | |
552 | BUG_ON(!(ctx->flags & SPU_CREATE_NOSCHED)); | |
553 | ||
4e0f4ed0 | 554 | spin_lock(&spu_prio->runq_lock); |
079cdb61 | 555 | prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE); |
a68cf983 | 556 | if (!signal_pending(current)) { |
4e0f4ed0 LB |
557 | __spu_add_to_rq(ctx); |
558 | spin_unlock(&spu_prio->runq_lock); | |
650f8b02 | 559 | mutex_unlock(&ctx->state_mutex); |
a68cf983 | 560 | schedule(); |
650f8b02 | 561 | mutex_lock(&ctx->state_mutex); |
4e0f4ed0 LB |
562 | spin_lock(&spu_prio->runq_lock); |
563 | __spu_del_from_rq(ctx); | |
8b3d6663 | 564 | } |
4e0f4ed0 | 565 | spin_unlock(&spu_prio->runq_lock); |
079cdb61 CH |
566 | __set_current_state(TASK_RUNNING); |
567 | remove_wait_queue(&ctx->stop_wq, &wait); | |
8b3d6663 AB |
568 | } |
569 | ||
079cdb61 | 570 | static struct spu *spu_get_idle(struct spu_context *ctx) |
a68cf983 | 571 | { |
36ddbb13 | 572 | struct spu *spu, *aff_ref_spu; |
486acd48 CH |
573 | int node, n; |
574 | ||
038200cf CH |
575 | spu_context_nospu_trace(spu_get_idle__enter, ctx); |
576 | ||
36ddbb13 AD |
577 | if (ctx->gang) { |
578 | mutex_lock(&ctx->gang->aff_mutex); | |
579 | if (has_affinity(ctx)) { | |
580 | aff_ref_spu = ctx->gang->aff_ref_spu; | |
581 | atomic_inc(&ctx->gang->aff_sched_count); | |
582 | mutex_unlock(&ctx->gang->aff_mutex); | |
583 | node = aff_ref_spu->node; | |
584 | ||
585 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
586 | spu = ctx_location(aff_ref_spu, ctx->aff_offset, node); | |
587 | if (spu && spu->alloc_state == SPU_FREE) | |
588 | goto found; | |
589 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
a68cf983 | 590 | |
0855b543 | 591 | atomic_dec(&ctx->gang->aff_sched_count); |
038200cf | 592 | goto not_found; |
36ddbb13 AD |
593 | } |
594 | mutex_unlock(&ctx->gang->aff_mutex); | |
595 | } | |
486acd48 | 596 | node = cpu_to_node(raw_smp_processor_id()); |
a68cf983 MN |
597 | for (n = 0; n < MAX_NUMNODES; n++, node++) { |
598 | node = (node < MAX_NUMNODES) ? node : 0; | |
ea1ae594 | 599 | if (!node_allowed(ctx, node)) |
a68cf983 | 600 | continue; |
486acd48 CH |
601 | |
602 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
603 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { | |
604 | if (spu->alloc_state == SPU_FREE) | |
605 | goto found; | |
606 | } | |
607 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
a68cf983 | 608 | } |
486acd48 | 609 | |
038200cf CH |
610 | not_found: |
611 | spu_context_nospu_trace(spu_get_idle__not_found, ctx); | |
486acd48 CH |
612 | return NULL; |
613 | ||
614 | found: | |
615 | spu->alloc_state = SPU_USED; | |
616 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
038200cf | 617 | spu_context_trace(spu_get_idle__found, ctx, spu); |
486acd48 | 618 | spu_init_channels(spu); |
a68cf983 MN |
619 | return spu; |
620 | } | |
8b3d6663 | 621 | |
52f04fcf CH |
622 | /** |
623 | * find_victim - find a lower priority context to preempt | |
624 | * @ctx: canidate context for running | |
625 | * | |
626 | * Returns the freed physical spu to run the new context on. | |
627 | */ | |
628 | static struct spu *find_victim(struct spu_context *ctx) | |
629 | { | |
630 | struct spu_context *victim = NULL; | |
631 | struct spu *spu; | |
632 | int node, n; | |
633 | ||
8a476d49 | 634 | spu_context_nospu_trace(spu_find_victim__enter, ctx); |
038200cf | 635 | |
52f04fcf CH |
636 | /* |
637 | * Look for a possible preemption candidate on the local node first. | |
638 | * If there is no candidate look at the other nodes. This isn't | |
9b1d21f8 | 639 | * exactly fair, but so far the whole spu scheduler tries to keep |
52f04fcf CH |
640 | * a strong node affinity. We might want to fine-tune this in |
641 | * the future. | |
642 | */ | |
643 | restart: | |
644 | node = cpu_to_node(raw_smp_processor_id()); | |
645 | for (n = 0; n < MAX_NUMNODES; n++, node++) { | |
646 | node = (node < MAX_NUMNODES) ? node : 0; | |
ea1ae594 | 647 | if (!node_allowed(ctx, node)) |
52f04fcf CH |
648 | continue; |
649 | ||
486acd48 CH |
650 | mutex_lock(&cbe_spu_info[node].list_mutex); |
651 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { | |
52f04fcf CH |
652 | struct spu_context *tmp = spu->ctx; |
653 | ||
c0e7b4aa | 654 | if (tmp && tmp->prio > ctx->prio && |
e65c2f6f | 655 | !(tmp->flags & SPU_CREATE_NOSCHED) && |
8d5636fb | 656 | (!victim || tmp->prio > victim->prio)) { |
52f04fcf | 657 | victim = spu->ctx; |
8d5636fb | 658 | } |
52f04fcf | 659 | } |
9f43e391 JK |
660 | if (victim) |
661 | get_spu_context(victim); | |
486acd48 | 662 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
52f04fcf CH |
663 | |
664 | if (victim) { | |
665 | /* | |
666 | * This nests ctx->state_mutex, but we always lock | |
667 | * higher priority contexts before lower priority | |
668 | * ones, so this is safe until we introduce | |
669 | * priority inheritance schemes. | |
91569531 LB |
670 | * |
671 | * XXX if the highest priority context is locked, | |
672 | * this can loop a long time. Might be better to | |
673 | * look at another context or give up after X retries. | |
52f04fcf CH |
674 | */ |
675 | if (!mutex_trylock(&victim->state_mutex)) { | |
8d5636fb | 676 | put_spu_context(victim); |
52f04fcf CH |
677 | victim = NULL; |
678 | goto restart; | |
679 | } | |
680 | ||
681 | spu = victim->spu; | |
b192541b | 682 | if (!spu || victim->prio <= ctx->prio) { |
52f04fcf CH |
683 | /* |
684 | * This race can happen because we've dropped | |
b192541b | 685 | * the active list mutex. Not a problem, just |
52f04fcf CH |
686 | * restart the search. |
687 | */ | |
688 | mutex_unlock(&victim->state_mutex); | |
8d5636fb | 689 | put_spu_context(victim); |
52f04fcf CH |
690 | victim = NULL; |
691 | goto restart; | |
692 | } | |
486acd48 | 693 | |
038200cf CH |
694 | spu_context_trace(__spu_deactivate__unload, ctx, spu); |
695 | ||
486acd48 CH |
696 | mutex_lock(&cbe_spu_info[node].list_mutex); |
697 | cbe_spu_info[node].nr_active--; | |
c0e7b4aa | 698 | spu_unbind_context(spu, victim); |
486acd48 CH |
699 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
700 | ||
e9f8a0b6 | 701 | victim->stats.invol_ctx_switch++; |
fe2f896d | 702 | spu->stats.invol_ctx_switch++; |
08fcf1d6 | 703 | if (test_bit(SPU_SCHED_SPU_RUN, &victim->sched_flags)) |
7a28a154 | 704 | spu_add_to_rq(victim); |
e65c2f6f | 705 | |
52f04fcf | 706 | mutex_unlock(&victim->state_mutex); |
8d5636fb | 707 | put_spu_context(victim); |
e65c2f6f | 708 | |
52f04fcf CH |
709 | return spu; |
710 | } | |
711 | } | |
712 | ||
713 | return NULL; | |
714 | } | |
715 | ||
e65c2f6f LB |
716 | static void __spu_schedule(struct spu *spu, struct spu_context *ctx) |
717 | { | |
718 | int node = spu->node; | |
719 | int success = 0; | |
720 | ||
721 | spu_set_timeslice(ctx); | |
722 | ||
723 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
724 | if (spu->ctx == NULL) { | |
725 | spu_bind_context(spu, ctx); | |
726 | cbe_spu_info[node].nr_active++; | |
727 | spu->alloc_state = SPU_USED; | |
728 | success = 1; | |
729 | } | |
730 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
731 | ||
732 | if (success) | |
733 | wake_up_all(&ctx->run_wq); | |
734 | else | |
735 | spu_add_to_rq(ctx); | |
736 | } | |
737 | ||
738 | static void spu_schedule(struct spu *spu, struct spu_context *ctx) | |
739 | { | |
c9101bdb CH |
740 | /* not a candidate for interruptible because it's called either |
741 | from the scheduler thread or from spu_deactivate */ | |
742 | mutex_lock(&ctx->state_mutex); | |
b2e601d1 AD |
743 | if (ctx->state == SPU_STATE_SAVED) |
744 | __spu_schedule(spu, ctx); | |
e65c2f6f LB |
745 | spu_release(ctx); |
746 | } | |
747 | ||
b65fe035 JK |
748 | /** |
749 | * spu_unschedule - remove a context from a spu, and possibly release it. | |
750 | * @spu: The SPU to unschedule from | |
751 | * @ctx: The context currently scheduled on the SPU | |
752 | * @free_spu Whether to free the SPU for other contexts | |
753 | * | |
754 | * Unbinds the context @ctx from the SPU @spu. If @free_spu is non-zero, the | |
755 | * SPU is made available for other contexts (ie, may be returned by | |
756 | * spu_get_idle). If this is zero, the caller is expected to schedule another | |
757 | * context to this spu. | |
758 | * | |
759 | * Should be called with ctx->state_mutex held. | |
760 | */ | |
761 | static void spu_unschedule(struct spu *spu, struct spu_context *ctx, | |
762 | int free_spu) | |
e65c2f6f LB |
763 | { |
764 | int node = spu->node; | |
765 | ||
766 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
767 | cbe_spu_info[node].nr_active--; | |
b65fe035 JK |
768 | if (free_spu) |
769 | spu->alloc_state = SPU_FREE; | |
e65c2f6f LB |
770 | spu_unbind_context(spu, ctx); |
771 | ctx->stats.invol_ctx_switch++; | |
772 | spu->stats.invol_ctx_switch++; | |
773 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
774 | } | |
775 | ||
079cdb61 CH |
776 | /** |
777 | * spu_activate - find a free spu for a context and execute it | |
778 | * @ctx: spu context to schedule | |
779 | * @flags: flags (currently ignored) | |
780 | * | |
08873095 | 781 | * Tries to find a free spu to run @ctx. If no free spu is available |
079cdb61 CH |
782 | * add the context to the runqueue so it gets woken up once an spu |
783 | * is available. | |
784 | */ | |
26bec673 | 785 | int spu_activate(struct spu_context *ctx, unsigned long flags) |
8b3d6663 | 786 | { |
e65c2f6f | 787 | struct spu *spu; |
079cdb61 | 788 | |
e65c2f6f LB |
789 | /* |
790 | * If there are multiple threads waiting for a single context | |
791 | * only one actually binds the context while the others will | |
792 | * only be able to acquire the state_mutex once the context | |
793 | * already is in runnable state. | |
794 | */ | |
795 | if (ctx->spu) | |
796 | return 0; | |
27449971 | 797 | |
e65c2f6f LB |
798 | spu_activate_top: |
799 | if (signal_pending(current)) | |
800 | return -ERESTARTSYS; | |
486acd48 | 801 | |
e65c2f6f LB |
802 | spu = spu_get_idle(ctx); |
803 | /* | |
804 | * If this is a realtime thread we try to get it running by | |
805 | * preempting a lower priority thread. | |
806 | */ | |
807 | if (!spu && rt_prio(ctx->prio)) | |
808 | spu = find_victim(ctx); | |
809 | if (spu) { | |
810 | unsigned long runcntl; | |
811 | ||
812 | runcntl = ctx->ops->runcntl_read(ctx); | |
813 | __spu_schedule(spu, ctx); | |
814 | if (runcntl & SPU_RUNCNTL_RUNNABLE) | |
815 | spuctx_switch_state(ctx, SPU_UTIL_USER); | |
079cdb61 | 816 | |
e65c2f6f LB |
817 | return 0; |
818 | } | |
819 | ||
820 | if (ctx->flags & SPU_CREATE_NOSCHED) { | |
50b520d4 | 821 | spu_prio_wait(ctx); |
e65c2f6f LB |
822 | goto spu_activate_top; |
823 | } | |
824 | ||
825 | spu_add_to_rq(ctx); | |
079cdb61 | 826 | |
e65c2f6f | 827 | return 0; |
8b3d6663 AB |
828 | } |
829 | ||
bb5db29a CH |
830 | /** |
831 | * grab_runnable_context - try to find a runnable context | |
832 | * | |
833 | * Remove the highest priority context on the runqueue and return it | |
834 | * to the caller. Returns %NULL if no runnable context was found. | |
835 | */ | |
ea1ae594 | 836 | static struct spu_context *grab_runnable_context(int prio, int node) |
bb5db29a | 837 | { |
ea1ae594 | 838 | struct spu_context *ctx; |
bb5db29a CH |
839 | int best; |
840 | ||
841 | spin_lock(&spu_prio->runq_lock); | |
7e90b749 | 842 | best = find_first_bit(spu_prio->bitmap, prio); |
ea1ae594 | 843 | while (best < prio) { |
bb5db29a CH |
844 | struct list_head *rq = &spu_prio->runq[best]; |
845 | ||
ea1ae594 CH |
846 | list_for_each_entry(ctx, rq, rq) { |
847 | /* XXX(hch): check for affinity here aswell */ | |
848 | if (__node_allowed(ctx, node)) { | |
849 | __spu_del_from_rq(ctx); | |
850 | goto found; | |
851 | } | |
852 | } | |
853 | best++; | |
bb5db29a | 854 | } |
ea1ae594 CH |
855 | ctx = NULL; |
856 | found: | |
bb5db29a | 857 | spin_unlock(&spu_prio->runq_lock); |
bb5db29a CH |
858 | return ctx; |
859 | } | |
860 | ||
861 | static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio) | |
862 | { | |
863 | struct spu *spu = ctx->spu; | |
864 | struct spu_context *new = NULL; | |
865 | ||
866 | if (spu) { | |
ea1ae594 | 867 | new = grab_runnable_context(max_prio, spu->node); |
bb5db29a | 868 | if (new || force) { |
b65fe035 | 869 | spu_unschedule(spu, ctx, new == NULL); |
e65c2f6f LB |
870 | if (new) { |
871 | if (new->flags & SPU_CREATE_NOSCHED) | |
872 | wake_up(&new->stop_wq); | |
873 | else { | |
874 | spu_release(ctx); | |
875 | spu_schedule(spu, new); | |
c9101bdb CH |
876 | /* this one can't easily be made |
877 | interruptible */ | |
878 | mutex_lock(&ctx->state_mutex); | |
e65c2f6f LB |
879 | } |
880 | } | |
bb5db29a | 881 | } |
bb5db29a CH |
882 | } |
883 | ||
884 | return new != NULL; | |
885 | } | |
886 | ||
678b2ff1 CH |
887 | /** |
888 | * spu_deactivate - unbind a context from it's physical spu | |
889 | * @ctx: spu context to unbind | |
890 | * | |
891 | * Unbind @ctx from the physical spu it is running on and schedule | |
892 | * the highest priority context to run on the freed physical spu. | |
893 | */ | |
8b3d6663 AB |
894 | void spu_deactivate(struct spu_context *ctx) |
895 | { | |
038200cf | 896 | spu_context_nospu_trace(spu_deactivate__enter, ctx); |
bb5db29a | 897 | __spu_deactivate(ctx, 1, MAX_PRIO); |
8b3d6663 AB |
898 | } |
899 | ||
ae7b4c52 | 900 | /** |
1474855d | 901 | * spu_yield - yield a physical spu if others are waiting |
ae7b4c52 CH |
902 | * @ctx: spu context to yield |
903 | * | |
904 | * Check if there is a higher priority context waiting and if yes | |
905 | * unbind @ctx from the physical spu and schedule the highest | |
906 | * priority context to run on the freed physical spu instead. | |
907 | */ | |
8b3d6663 AB |
908 | void spu_yield(struct spu_context *ctx) |
909 | { | |
038200cf | 910 | spu_context_nospu_trace(spu_yield__enter, ctx); |
e5c0b9ec CH |
911 | if (!(ctx->flags & SPU_CREATE_NOSCHED)) { |
912 | mutex_lock(&ctx->state_mutex); | |
27ec41d3 | 913 | __spu_deactivate(ctx, 0, MAX_PRIO); |
e5c0b9ec CH |
914 | mutex_unlock(&ctx->state_mutex); |
915 | } | |
bb5db29a | 916 | } |
8b3d6663 | 917 | |
486acd48 | 918 | static noinline void spusched_tick(struct spu_context *ctx) |
bb5db29a | 919 | { |
e65c2f6f LB |
920 | struct spu_context *new = NULL; |
921 | struct spu *spu = NULL; | |
e65c2f6f | 922 | |
c9101bdb CH |
923 | if (spu_acquire(ctx)) |
924 | BUG(); /* a kernel thread never has signals pending */ | |
e65c2f6f LB |
925 | |
926 | if (ctx->state != SPU_STATE_RUNNABLE) | |
927 | goto out; | |
df09cf3e | 928 | if (ctx->flags & SPU_CREATE_NOSCHED) |
e65c2f6f | 929 | goto out; |
df09cf3e | 930 | if (ctx->policy == SCHED_FIFO) |
e65c2f6f | 931 | goto out; |
df09cf3e | 932 | |
ce7c191b | 933 | if (--ctx->time_slice && test_bit(SPU_SCHED_SPU_RUN, &ctx->sched_flags)) |
e65c2f6f | 934 | goto out; |
bb5db29a | 935 | |
e65c2f6f | 936 | spu = ctx->spu; |
038200cf CH |
937 | |
938 | spu_context_trace(spusched_tick__preempt, ctx, spu); | |
939 | ||
e65c2f6f LB |
940 | new = grab_runnable_context(ctx->prio + 1, spu->node); |
941 | if (new) { | |
b65fe035 | 942 | spu_unschedule(spu, ctx, 0); |
ce7c191b | 943 | if (test_bit(SPU_SCHED_SPU_RUN, &ctx->sched_flags)) |
4ef11014 | 944 | spu_add_to_rq(ctx); |
bb5db29a | 945 | } else { |
038200cf | 946 | spu_context_nospu_trace(spusched_tick__newslice, ctx); |
2442a8ba LB |
947 | if (!ctx->time_slice) |
948 | ctx->time_slice++; | |
8b3d6663 | 949 | } |
e65c2f6f LB |
950 | out: |
951 | spu_release(ctx); | |
952 | ||
953 | if (new) | |
954 | spu_schedule(spu, new); | |
8b3d6663 AB |
955 | } |
956 | ||
65de66f0 CH |
957 | /** |
958 | * count_active_contexts - count nr of active tasks | |
959 | * | |
960 | * Return the number of tasks currently running or waiting to run. | |
961 | * | |
486acd48 | 962 | * Note that we don't take runq_lock / list_mutex here. Reading |
65de66f0 CH |
963 | * a single 32bit value is atomic on powerpc, and we don't care |
964 | * about memory ordering issues here. | |
965 | */ | |
966 | static unsigned long count_active_contexts(void) | |
967 | { | |
968 | int nr_active = 0, node; | |
969 | ||
970 | for (node = 0; node < MAX_NUMNODES; node++) | |
486acd48 | 971 | nr_active += cbe_spu_info[node].nr_active; |
65de66f0 CH |
972 | nr_active += spu_prio->nr_waiting; |
973 | ||
974 | return nr_active; | |
975 | } | |
976 | ||
977 | /** | |
90608a29 | 978 | * spu_calc_load - update the avenrun load estimates. |
65de66f0 CH |
979 | * |
980 | * No locking against reading these values from userspace, as for | |
981 | * the CPU loadavg code. | |
982 | */ | |
90608a29 | 983 | static void spu_calc_load(void) |
65de66f0 CH |
984 | { |
985 | unsigned long active_tasks; /* fixed-point */ | |
90608a29 AL |
986 | |
987 | active_tasks = count_active_contexts() * FIXED_1; | |
988 | CALC_LOAD(spu_avenrun[0], EXP_1, active_tasks); | |
989 | CALC_LOAD(spu_avenrun[1], EXP_5, active_tasks); | |
990 | CALC_LOAD(spu_avenrun[2], EXP_15, active_tasks); | |
65de66f0 CH |
991 | } |
992 | ||
37901802 CH |
993 | static void spusched_wake(unsigned long data) |
994 | { | |
995 | mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK); | |
996 | wake_up_process(spusched_task); | |
90608a29 AL |
997 | } |
998 | ||
999 | static void spuloadavg_wake(unsigned long data) | |
1000 | { | |
1001 | mod_timer(&spuloadavg_timer, jiffies + LOAD_FREQ); | |
1002 | spu_calc_load(); | |
37901802 CH |
1003 | } |
1004 | ||
1005 | static int spusched_thread(void *unused) | |
1006 | { | |
486acd48 | 1007 | struct spu *spu; |
37901802 CH |
1008 | int node; |
1009 | ||
37901802 CH |
1010 | while (!kthread_should_stop()) { |
1011 | set_current_state(TASK_INTERRUPTIBLE); | |
1012 | schedule(); | |
1013 | for (node = 0; node < MAX_NUMNODES; node++) { | |
e65c2f6f LB |
1014 | struct mutex *mtx = &cbe_spu_info[node].list_mutex; |
1015 | ||
1016 | mutex_lock(mtx); | |
1017 | list_for_each_entry(spu, &cbe_spu_info[node].spus, | |
1018 | cbe_list) { | |
1019 | struct spu_context *ctx = spu->ctx; | |
1020 | ||
1021 | if (ctx) { | |
8d5636fb | 1022 | get_spu_context(ctx); |
e65c2f6f LB |
1023 | mutex_unlock(mtx); |
1024 | spusched_tick(ctx); | |
1025 | mutex_lock(mtx); | |
8d5636fb | 1026 | put_spu_context(ctx); |
e65c2f6f LB |
1027 | } |
1028 | } | |
1029 | mutex_unlock(mtx); | |
37901802 CH |
1030 | } |
1031 | } | |
1032 | ||
37901802 CH |
1033 | return 0; |
1034 | } | |
1035 | ||
7cd58e43 JK |
1036 | void spuctx_switch_state(struct spu_context *ctx, |
1037 | enum spu_utilization_state new_state) | |
1038 | { | |
1039 | unsigned long long curtime; | |
1040 | signed long long delta; | |
1041 | struct timespec ts; | |
1042 | struct spu *spu; | |
1043 | enum spu_utilization_state old_state; | |
fabb6570 | 1044 | int node; |
7cd58e43 JK |
1045 | |
1046 | ktime_get_ts(&ts); | |
1047 | curtime = timespec_to_ns(&ts); | |
1048 | delta = curtime - ctx->stats.tstamp; | |
1049 | ||
1050 | WARN_ON(!mutex_is_locked(&ctx->state_mutex)); | |
1051 | WARN_ON(delta < 0); | |
1052 | ||
1053 | spu = ctx->spu; | |
1054 | old_state = ctx->stats.util_state; | |
1055 | ctx->stats.util_state = new_state; | |
1056 | ctx->stats.tstamp = curtime; | |
1057 | ||
1058 | /* | |
1059 | * Update the physical SPU utilization statistics. | |
1060 | */ | |
1061 | if (spu) { | |
1062 | ctx->stats.times[old_state] += delta; | |
1063 | spu->stats.times[old_state] += delta; | |
1064 | spu->stats.util_state = new_state; | |
1065 | spu->stats.tstamp = curtime; | |
fabb6570 MS |
1066 | node = spu->node; |
1067 | if (old_state == SPU_UTIL_USER) | |
1068 | atomic_dec(&cbe_spu_info[node].busy_spus); | |
cb9808d3 | 1069 | if (new_state == SPU_UTIL_USER) |
fabb6570 | 1070 | atomic_inc(&cbe_spu_info[node].busy_spus); |
7cd58e43 JK |
1071 | } |
1072 | } | |
1073 | ||
65de66f0 CH |
1074 | #define LOAD_INT(x) ((x) >> FSHIFT) |
1075 | #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100) | |
1076 | ||
1077 | static int show_spu_loadavg(struct seq_file *s, void *private) | |
1078 | { | |
1079 | int a, b, c; | |
1080 | ||
1081 | a = spu_avenrun[0] + (FIXED_1/200); | |
1082 | b = spu_avenrun[1] + (FIXED_1/200); | |
1083 | c = spu_avenrun[2] + (FIXED_1/200); | |
1084 | ||
1085 | /* | |
1086 | * Note that last_pid doesn't really make much sense for the | |
9b1d21f8 | 1087 | * SPU loadavg (it even seems very odd on the CPU side...), |
65de66f0 CH |
1088 | * but we include it here to have a 100% compatible interface. |
1089 | */ | |
1090 | seq_printf(s, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n", | |
1091 | LOAD_INT(a), LOAD_FRAC(a), | |
1092 | LOAD_INT(b), LOAD_FRAC(b), | |
1093 | LOAD_INT(c), LOAD_FRAC(c), | |
1094 | count_active_contexts(), | |
1095 | atomic_read(&nr_spu_contexts), | |
1096 | current->nsproxy->pid_ns->last_pid); | |
1097 | return 0; | |
1098 | } | |
1099 | ||
1100 | static int spu_loadavg_open(struct inode *inode, struct file *file) | |
1101 | { | |
1102 | return single_open(file, show_spu_loadavg, NULL); | |
1103 | } | |
1104 | ||
1105 | static const struct file_operations spu_loadavg_fops = { | |
1106 | .open = spu_loadavg_open, | |
1107 | .read = seq_read, | |
1108 | .llseek = seq_lseek, | |
1109 | .release = single_release, | |
1110 | }; | |
1111 | ||
8b3d6663 AB |
1112 | int __init spu_sched_init(void) |
1113 | { | |
65de66f0 CH |
1114 | struct proc_dir_entry *entry; |
1115 | int err = -ENOMEM, i; | |
8b3d6663 | 1116 | |
a68cf983 | 1117 | spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL); |
37901802 | 1118 | if (!spu_prio) |
65de66f0 | 1119 | goto out; |
37901802 | 1120 | |
8b3d6663 | 1121 | for (i = 0; i < MAX_PRIO; i++) { |
079cdb61 | 1122 | INIT_LIST_HEAD(&spu_prio->runq[i]); |
a68cf983 | 1123 | __clear_bit(i, spu_prio->bitmap); |
8b3d6663 | 1124 | } |
079cdb61 | 1125 | spin_lock_init(&spu_prio->runq_lock); |
37901802 | 1126 | |
c77239b8 | 1127 | setup_timer(&spusched_timer, spusched_wake, 0); |
90608a29 | 1128 | setup_timer(&spuloadavg_timer, spuloadavg_wake, 0); |
c77239b8 | 1129 | |
37901802 CH |
1130 | spusched_task = kthread_run(spusched_thread, NULL, "spusched"); |
1131 | if (IS_ERR(spusched_task)) { | |
65de66f0 CH |
1132 | err = PTR_ERR(spusched_task); |
1133 | goto out_free_spu_prio; | |
37901802 | 1134 | } |
f3f59bec | 1135 | |
90608a29 AL |
1136 | mod_timer(&spuloadavg_timer, 0); |
1137 | ||
66747138 | 1138 | entry = proc_create("spu_loadavg", 0, NULL, &spu_loadavg_fops); |
65de66f0 CH |
1139 | if (!entry) |
1140 | goto out_stop_kthread; | |
65de66f0 | 1141 | |
f3f59bec JK |
1142 | pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n", |
1143 | SPUSCHED_TICK, MIN_SPU_TIMESLICE, DEF_SPU_TIMESLICE); | |
8b3d6663 | 1144 | return 0; |
37901802 | 1145 | |
65de66f0 CH |
1146 | out_stop_kthread: |
1147 | kthread_stop(spusched_task); | |
1148 | out_free_spu_prio: | |
1149 | kfree(spu_prio); | |
1150 | out: | |
1151 | return err; | |
8b3d6663 AB |
1152 | } |
1153 | ||
d1450317 | 1154 | void spu_sched_exit(void) |
8b3d6663 | 1155 | { |
486acd48 | 1156 | struct spu *spu; |
a68cf983 MN |
1157 | int node; |
1158 | ||
65de66f0 CH |
1159 | remove_proc_entry("spu_loadavg", NULL); |
1160 | ||
c77239b8 | 1161 | del_timer_sync(&spusched_timer); |
90608a29 | 1162 | del_timer_sync(&spuloadavg_timer); |
37901802 CH |
1163 | kthread_stop(spusched_task); |
1164 | ||
a68cf983 | 1165 | for (node = 0; node < MAX_NUMNODES; node++) { |
486acd48 CH |
1166 | mutex_lock(&cbe_spu_info[node].list_mutex); |
1167 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) | |
1168 | if (spu->alloc_state != SPU_FREE) | |
1169 | spu->alloc_state = SPU_FREE; | |
1170 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
8b3d6663 | 1171 | } |
a68cf983 | 1172 | kfree(spu_prio); |
8b3d6663 | 1173 | } |