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