]>
Commit | Line | Data |
---|---|---|
8b3d6663 AB |
1 | /* sched.c - SPU scheduler. |
2 | * | |
3 | * Copyright (C) IBM 2005 | |
4 | * Author: Mark Nutter <mnutter@us.ibm.com> | |
5 | * | |
a68cf983 | 6 | * 2006-03-31 NUMA domains added. |
8b3d6663 AB |
7 | * |
8 | * This program is free software; you can redistribute it and/or modify | |
9 | * it under the terms of the GNU General Public License as published by | |
10 | * the Free Software Foundation; either version 2, or (at your option) | |
11 | * any later version. | |
12 | * | |
13 | * This program is distributed in the hope that it will be useful, | |
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
16 | * GNU General Public License for more details. | |
17 | * | |
18 | * You should have received a copy of the GNU General Public License | |
19 | * along with this program; if not, write to the Free Software | |
20 | * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | |
21 | */ | |
22 | ||
3b3d22cb AB |
23 | #undef DEBUG |
24 | ||
8b3d6663 AB |
25 | #include <linux/module.h> |
26 | #include <linux/errno.h> | |
27 | #include <linux/sched.h> | |
28 | #include <linux/kernel.h> | |
29 | #include <linux/mm.h> | |
30 | #include <linux/completion.h> | |
31 | #include <linux/vmalloc.h> | |
32 | #include <linux/smp.h> | |
8b3d6663 AB |
33 | #include <linux/stddef.h> |
34 | #include <linux/unistd.h> | |
a68cf983 MN |
35 | #include <linux/numa.h> |
36 | #include <linux/mutex.h> | |
86767277 | 37 | #include <linux/notifier.h> |
37901802 | 38 | #include <linux/kthread.h> |
65de66f0 CH |
39 | #include <linux/pid_namespace.h> |
40 | #include <linux/proc_fs.h> | |
41 | #include <linux/seq_file.h> | |
8b3d6663 AB |
42 | |
43 | #include <asm/io.h> | |
44 | #include <asm/mmu_context.h> | |
45 | #include <asm/spu.h> | |
46 | #include <asm/spu_csa.h> | |
a91942ae | 47 | #include <asm/spu_priv1.h> |
8b3d6663 AB |
48 | #include "spufs.h" |
49 | ||
8b3d6663 | 50 | struct spu_prio_array { |
72cb3608 | 51 | DECLARE_BITMAP(bitmap, MAX_PRIO); |
079cdb61 CH |
52 | struct list_head runq[MAX_PRIO]; |
53 | spinlock_t runq_lock; | |
a68cf983 MN |
54 | struct list_head active_list[MAX_NUMNODES]; |
55 | struct mutex active_mutex[MAX_NUMNODES]; | |
65de66f0 CH |
56 | int nr_active[MAX_NUMNODES]; |
57 | int nr_waiting; | |
8b3d6663 AB |
58 | }; |
59 | ||
65de66f0 | 60 | static unsigned long spu_avenrun[3]; |
a68cf983 | 61 | static struct spu_prio_array *spu_prio; |
37901802 CH |
62 | static struct task_struct *spusched_task; |
63 | static struct timer_list spusched_timer; | |
8b3d6663 | 64 | |
fe443ef2 CH |
65 | /* |
66 | * Priority of a normal, non-rt, non-niced'd process (aka nice level 0). | |
67 | */ | |
68 | #define NORMAL_PRIO 120 | |
69 | ||
70 | /* | |
71 | * Frequency of the spu scheduler tick. By default we do one SPU scheduler | |
72 | * tick for every 10 CPU scheduler ticks. | |
73 | */ | |
74 | #define SPUSCHED_TICK (10) | |
75 | ||
76 | /* | |
77 | * These are the 'tuning knobs' of the scheduler: | |
78 | * | |
60e24239 JK |
79 | * Minimum timeslice is 5 msecs (or 1 spu scheduler tick, whichever is |
80 | * larger), default timeslice is 100 msecs, maximum timeslice is 800 msecs. | |
fe443ef2 | 81 | */ |
60e24239 JK |
82 | #define MIN_SPU_TIMESLICE max(5 * HZ / (1000 * SPUSCHED_TICK), 1) |
83 | #define DEF_SPU_TIMESLICE (100 * HZ / (1000 * SPUSCHED_TICK)) | |
fe443ef2 CH |
84 | |
85 | #define MAX_USER_PRIO (MAX_PRIO - MAX_RT_PRIO) | |
86 | #define SCALE_PRIO(x, prio) \ | |
87 | max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_SPU_TIMESLICE) | |
88 | ||
89 | /* | |
90 | * scale user-nice values [ -20 ... 0 ... 19 ] to time slice values: | |
91 | * [800ms ... 100ms ... 5ms] | |
92 | * | |
93 | * The higher a thread's priority, the bigger timeslices | |
94 | * it gets during one round of execution. But even the lowest | |
95 | * priority thread gets MIN_TIMESLICE worth of execution time. | |
96 | */ | |
97 | void spu_set_timeslice(struct spu_context *ctx) | |
98 | { | |
99 | if (ctx->prio < NORMAL_PRIO) | |
100 | ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE * 4, ctx->prio); | |
101 | else | |
102 | ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE, ctx->prio); | |
103 | } | |
104 | ||
2cf2b3b4 CH |
105 | /* |
106 | * Update scheduling information from the owning thread. | |
107 | */ | |
108 | void __spu_update_sched_info(struct spu_context *ctx) | |
109 | { | |
476273ad CH |
110 | /* |
111 | * 32-Bit assignment are atomic on powerpc, and we don't care about | |
112 | * memory ordering here because retriving the controlling thread is | |
113 | * per defintion racy. | |
114 | */ | |
115 | ctx->tid = current->pid; | |
116 | ||
2cf2b3b4 CH |
117 | /* |
118 | * We do our own priority calculations, so we normally want | |
119 | * ->static_prio to start with. Unfortunately thies field | |
120 | * contains junk for threads with a realtime scheduling | |
121 | * policy so we have to look at ->prio in this case. | |
122 | */ | |
123 | if (rt_prio(current->prio)) | |
124 | ctx->prio = current->prio; | |
125 | else | |
126 | ctx->prio = current->static_prio; | |
127 | ctx->policy = current->policy; | |
ea1ae594 CH |
128 | |
129 | /* | |
130 | * A lot of places that don't hold active_mutex poke into | |
131 | * cpus_allowed, including grab_runnable_context which | |
132 | * already holds the runq_lock. So abuse runq_lock | |
133 | * to protect this field aswell. | |
134 | */ | |
135 | spin_lock(&spu_prio->runq_lock); | |
136 | ctx->cpus_allowed = current->cpus_allowed; | |
137 | spin_unlock(&spu_prio->runq_lock); | |
2cf2b3b4 CH |
138 | } |
139 | ||
140 | void spu_update_sched_info(struct spu_context *ctx) | |
141 | { | |
142 | int node = ctx->spu->node; | |
143 | ||
144 | mutex_lock(&spu_prio->active_mutex[node]); | |
145 | __spu_update_sched_info(ctx); | |
146 | mutex_unlock(&spu_prio->active_mutex[node]); | |
147 | } | |
148 | ||
ea1ae594 | 149 | static int __node_allowed(struct spu_context *ctx, int node) |
8b3d6663 | 150 | { |
ea1ae594 CH |
151 | if (nr_cpus_node(node)) { |
152 | cpumask_t mask = node_to_cpumask(node); | |
8b3d6663 | 153 | |
ea1ae594 CH |
154 | if (cpus_intersects(mask, ctx->cpus_allowed)) |
155 | return 1; | |
156 | } | |
157 | ||
158 | return 0; | |
159 | } | |
160 | ||
161 | static int node_allowed(struct spu_context *ctx, int node) | |
162 | { | |
163 | int rval; | |
164 | ||
165 | spin_lock(&spu_prio->runq_lock); | |
166 | rval = __node_allowed(ctx, node); | |
167 | spin_unlock(&spu_prio->runq_lock); | |
168 | ||
169 | return rval; | |
8b3d6663 AB |
170 | } |
171 | ||
202557d2 CH |
172 | /** |
173 | * spu_add_to_active_list - add spu to active list | |
174 | * @spu: spu to add to the active list | |
175 | */ | |
176 | static void spu_add_to_active_list(struct spu *spu) | |
177 | { | |
65de66f0 CH |
178 | int node = spu->node; |
179 | ||
180 | mutex_lock(&spu_prio->active_mutex[node]); | |
181 | spu_prio->nr_active[node]++; | |
182 | list_add_tail(&spu->list, &spu_prio->active_list[node]); | |
183 | mutex_unlock(&spu_prio->active_mutex[node]); | |
202557d2 CH |
184 | } |
185 | ||
37901802 CH |
186 | static void __spu_remove_from_active_list(struct spu *spu) |
187 | { | |
188 | list_del_init(&spu->list); | |
65de66f0 | 189 | spu_prio->nr_active[spu->node]--; |
37901802 CH |
190 | } |
191 | ||
202557d2 CH |
192 | /** |
193 | * spu_remove_from_active_list - remove spu from active list | |
194 | * @spu: spu to remove from the active list | |
202557d2 | 195 | */ |
678b2ff1 | 196 | static void spu_remove_from_active_list(struct spu *spu) |
202557d2 CH |
197 | { |
198 | int node = spu->node; | |
202557d2 CH |
199 | |
200 | mutex_lock(&spu_prio->active_mutex[node]); | |
37901802 | 201 | __spu_remove_from_active_list(spu); |
202557d2 | 202 | mutex_unlock(&spu_prio->active_mutex[node]); |
202557d2 CH |
203 | } |
204 | ||
86767277 AB |
205 | static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier); |
206 | ||
207 | static void spu_switch_notify(struct spu *spu, struct spu_context *ctx) | |
208 | { | |
209 | blocking_notifier_call_chain(&spu_switch_notifier, | |
210 | ctx ? ctx->object_id : 0, spu); | |
211 | } | |
212 | ||
213 | int spu_switch_event_register(struct notifier_block * n) | |
214 | { | |
215 | return blocking_notifier_chain_register(&spu_switch_notifier, n); | |
216 | } | |
217 | ||
218 | int spu_switch_event_unregister(struct notifier_block * n) | |
219 | { | |
220 | return blocking_notifier_chain_unregister(&spu_switch_notifier, n); | |
221 | } | |
222 | ||
202557d2 CH |
223 | /** |
224 | * spu_bind_context - bind spu context to physical spu | |
225 | * @spu: physical spu to bind to | |
226 | * @ctx: context to bind | |
227 | */ | |
228 | static void spu_bind_context(struct spu *spu, struct spu_context *ctx) | |
8b3d6663 | 229 | { |
a68cf983 MN |
230 | pr_debug("%s: pid=%d SPU=%d NODE=%d\n", __FUNCTION__, current->pid, |
231 | spu->number, spu->node); | |
27ec41d3 | 232 | spuctx_switch_state(ctx, SPU_UTIL_SYSTEM); |
e9f8a0b6 | 233 | |
aa6d5b20 AB |
234 | if (ctx->flags & SPU_CREATE_NOSCHED) |
235 | atomic_inc(&cbe_spu_info[spu->node].reserved_spus); | |
c5fc8d2a AB |
236 | if (!list_empty(&ctx->aff_list)) |
237 | atomic_inc(&ctx->gang->aff_sched_count); | |
aa6d5b20 | 238 | |
e9f8a0b6 CH |
239 | ctx->stats.slb_flt_base = spu->stats.slb_flt; |
240 | ctx->stats.class2_intr_base = spu->stats.class2_intr; | |
241 | ||
8b3d6663 AB |
242 | spu->ctx = ctx; |
243 | spu->flags = 0; | |
244 | ctx->spu = spu; | |
245 | ctx->ops = &spu_hw_ops; | |
246 | spu->pid = current->pid; | |
94b2a439 | 247 | spu_associate_mm(spu, ctx->owner); |
8b3d6663 AB |
248 | spu->ibox_callback = spufs_ibox_callback; |
249 | spu->wbox_callback = spufs_wbox_callback; | |
5110459f | 250 | spu->stop_callback = spufs_stop_callback; |
a33a7d73 | 251 | spu->mfc_callback = spufs_mfc_callback; |
9add11da | 252 | spu->dma_callback = spufs_dma_callback; |
8b3d6663 | 253 | mb(); |
5110459f | 254 | spu_unmap_mappings(ctx); |
8b3d6663 | 255 | spu_restore(&ctx->csa, spu); |
2a911f0b | 256 | spu->timestamp = jiffies; |
a68cf983 | 257 | spu_cpu_affinity_set(spu, raw_smp_processor_id()); |
86767277 | 258 | spu_switch_notify(spu, ctx); |
81998baf | 259 | ctx->state = SPU_STATE_RUNNABLE; |
27ec41d3 AD |
260 | |
261 | spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED); | |
8b3d6663 AB |
262 | } |
263 | ||
c5fc8d2a AB |
264 | /* |
265 | * XXX(hch): needs locking. | |
266 | */ | |
267 | static inline int sched_spu(struct spu *spu) | |
268 | { | |
269 | return (!spu->ctx || !(spu->ctx->flags & SPU_CREATE_NOSCHED)); | |
270 | } | |
271 | ||
272 | static void aff_merge_remaining_ctxs(struct spu_gang *gang) | |
273 | { | |
274 | struct spu_context *ctx; | |
275 | ||
276 | list_for_each_entry(ctx, &gang->aff_list_head, aff_list) { | |
277 | if (list_empty(&ctx->aff_list)) | |
278 | list_add(&ctx->aff_list, &gang->aff_list_head); | |
279 | } | |
280 | gang->aff_flags |= AFF_MERGED; | |
281 | } | |
282 | ||
283 | static void aff_set_offsets(struct spu_gang *gang) | |
284 | { | |
285 | struct spu_context *ctx; | |
286 | int offset; | |
287 | ||
288 | offset = -1; | |
289 | list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list, | |
290 | aff_list) { | |
291 | if (&ctx->aff_list == &gang->aff_list_head) | |
292 | break; | |
293 | ctx->aff_offset = offset--; | |
294 | } | |
295 | ||
296 | offset = 0; | |
297 | list_for_each_entry(ctx, gang->aff_ref_ctx->aff_list.prev, aff_list) { | |
298 | if (&ctx->aff_list == &gang->aff_list_head) | |
299 | break; | |
300 | ctx->aff_offset = offset++; | |
301 | } | |
302 | ||
303 | gang->aff_flags |= AFF_OFFSETS_SET; | |
304 | } | |
305 | ||
306 | static struct spu *aff_ref_location(struct spu_context *ctx, int mem_aff, | |
307 | int group_size, int lowest_offset) | |
308 | { | |
309 | struct spu *spu; | |
310 | int node, n; | |
311 | ||
312 | /* | |
313 | * TODO: A better algorithm could be used to find a good spu to be | |
314 | * used as reference location for the ctxs chain. | |
315 | */ | |
316 | node = cpu_to_node(raw_smp_processor_id()); | |
317 | for (n = 0; n < MAX_NUMNODES; n++, node++) { | |
318 | node = (node < MAX_NUMNODES) ? node : 0; | |
319 | if (!node_allowed(ctx, node)) | |
320 | continue; | |
321 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { | |
322 | if ((!mem_aff || spu->has_mem_affinity) && | |
323 | sched_spu(spu)) | |
324 | return spu; | |
325 | } | |
326 | } | |
327 | return NULL; | |
328 | } | |
329 | ||
330 | static void aff_set_ref_point_location(struct spu_gang *gang) | |
331 | { | |
332 | int mem_aff, gs, lowest_offset; | |
333 | struct spu_context *ctx; | |
334 | struct spu *tmp; | |
335 | ||
336 | mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM; | |
337 | lowest_offset = 0; | |
338 | gs = 0; | |
339 | ||
340 | list_for_each_entry(tmp, &gang->aff_list_head, aff_list) | |
341 | gs++; | |
342 | ||
343 | list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list, | |
344 | aff_list) { | |
345 | if (&ctx->aff_list == &gang->aff_list_head) | |
346 | break; | |
347 | lowest_offset = ctx->aff_offset; | |
348 | } | |
349 | ||
350 | gang->aff_ref_spu = aff_ref_location(ctx, mem_aff, gs, lowest_offset); | |
351 | } | |
352 | ||
353 | static struct spu *ctx_location(struct spu *ref, int offset) | |
354 | { | |
355 | struct spu *spu; | |
356 | ||
357 | spu = NULL; | |
358 | if (offset >= 0) { | |
359 | list_for_each_entry(spu, ref->aff_list.prev, aff_list) { | |
360 | if (offset == 0) | |
361 | break; | |
362 | if (sched_spu(spu)) | |
363 | offset--; | |
364 | } | |
365 | } else { | |
366 | list_for_each_entry_reverse(spu, ref->aff_list.next, aff_list) { | |
367 | if (offset == 0) | |
368 | break; | |
369 | if (sched_spu(spu)) | |
370 | offset++; | |
371 | } | |
372 | } | |
373 | return spu; | |
374 | } | |
375 | ||
376 | /* | |
377 | * affinity_check is called each time a context is going to be scheduled. | |
378 | * It returns the spu ptr on which the context must run. | |
379 | */ | |
380 | struct spu *affinity_check(struct spu_context *ctx) | |
381 | { | |
382 | struct spu_gang *gang; | |
383 | ||
384 | if (list_empty(&ctx->aff_list)) | |
385 | return NULL; | |
386 | gang = ctx->gang; | |
387 | mutex_lock(&gang->aff_mutex); | |
388 | if (!gang->aff_ref_spu) { | |
389 | if (!(gang->aff_flags & AFF_MERGED)) | |
390 | aff_merge_remaining_ctxs(gang); | |
391 | if (!(gang->aff_flags & AFF_OFFSETS_SET)) | |
392 | aff_set_offsets(gang); | |
393 | aff_set_ref_point_location(gang); | |
394 | } | |
395 | mutex_unlock(&gang->aff_mutex); | |
396 | if (!gang->aff_ref_spu) | |
397 | return NULL; | |
398 | return ctx_location(gang->aff_ref_spu, ctx->aff_offset); | |
399 | } | |
400 | ||
202557d2 CH |
401 | /** |
402 | * spu_unbind_context - unbind spu context from physical spu | |
403 | * @spu: physical spu to unbind from | |
404 | * @ctx: context to unbind | |
202557d2 | 405 | */ |
678b2ff1 | 406 | static void spu_unbind_context(struct spu *spu, struct spu_context *ctx) |
8b3d6663 | 407 | { |
a68cf983 MN |
408 | pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__, |
409 | spu->pid, spu->number, spu->node); | |
27ec41d3 | 410 | spuctx_switch_state(ctx, SPU_UTIL_SYSTEM); |
fe2f896d | 411 | |
aa6d5b20 AB |
412 | if (spu->ctx->flags & SPU_CREATE_NOSCHED) |
413 | atomic_dec(&cbe_spu_info[spu->node].reserved_spus); | |
c5fc8d2a AB |
414 | if (!list_empty(&ctx->aff_list)) |
415 | if (atomic_dec_and_test(&ctx->gang->aff_sched_count)) | |
416 | ctx->gang->aff_ref_spu = NULL; | |
86767277 | 417 | spu_switch_notify(spu, NULL); |
5110459f | 418 | spu_unmap_mappings(ctx); |
8b3d6663 | 419 | spu_save(&ctx->csa, spu); |
2a911f0b | 420 | spu->timestamp = jiffies; |
8b3d6663 AB |
421 | ctx->state = SPU_STATE_SAVED; |
422 | spu->ibox_callback = NULL; | |
423 | spu->wbox_callback = NULL; | |
5110459f | 424 | spu->stop_callback = NULL; |
a33a7d73 | 425 | spu->mfc_callback = NULL; |
9add11da | 426 | spu->dma_callback = NULL; |
94b2a439 | 427 | spu_associate_mm(spu, NULL); |
8b3d6663 | 428 | spu->pid = 0; |
8b3d6663 | 429 | ctx->ops = &spu_backing_ops; |
2a911f0b | 430 | spu->flags = 0; |
8b3d6663 | 431 | spu->ctx = NULL; |
e9f8a0b6 CH |
432 | |
433 | ctx->stats.slb_flt += | |
434 | (spu->stats.slb_flt - ctx->stats.slb_flt_base); | |
435 | ctx->stats.class2_intr += | |
436 | (spu->stats.class2_intr - ctx->stats.class2_intr_base); | |
27ec41d3 AD |
437 | |
438 | /* This maps the underlying spu state to idle */ | |
439 | spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED); | |
440 | ctx->spu = NULL; | |
8b3d6663 AB |
441 | } |
442 | ||
079cdb61 CH |
443 | /** |
444 | * spu_add_to_rq - add a context to the runqueue | |
445 | * @ctx: context to add | |
446 | */ | |
4e0f4ed0 | 447 | static void __spu_add_to_rq(struct spu_context *ctx) |
8b3d6663 | 448 | { |
27449971 CH |
449 | /* |
450 | * Unfortunately this code path can be called from multiple threads | |
451 | * on behalf of a single context due to the way the problem state | |
452 | * mmap support works. | |
453 | * | |
454 | * Fortunately we need to wake up all these threads at the same time | |
455 | * and can simply skip the runqueue addition for every but the first | |
456 | * thread getting into this codepath. | |
457 | * | |
458 | * It's still quite hacky, and long-term we should proxy all other | |
459 | * threads through the owner thread so that spu_run is in control | |
460 | * of all the scheduling activity for a given context. | |
461 | */ | |
462 | if (list_empty(&ctx->rq)) { | |
463 | list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]); | |
464 | set_bit(ctx->prio, spu_prio->bitmap); | |
465 | if (!spu_prio->nr_waiting++) | |
466 | __mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK); | |
467 | } | |
2a911f0b | 468 | } |
5110459f | 469 | |
4e0f4ed0 | 470 | static void __spu_del_from_rq(struct spu_context *ctx) |
a475c2f4 | 471 | { |
4e0f4ed0 LB |
472 | int prio = ctx->prio; |
473 | ||
65de66f0 | 474 | if (!list_empty(&ctx->rq)) { |
c77239b8 CH |
475 | if (!--spu_prio->nr_waiting) |
476 | del_timer(&spusched_timer); | |
a475c2f4 | 477 | list_del_init(&ctx->rq); |
c77239b8 CH |
478 | |
479 | if (list_empty(&spu_prio->runq[prio])) | |
480 | clear_bit(prio, spu_prio->bitmap); | |
65de66f0 | 481 | } |
079cdb61 | 482 | } |
a68cf983 | 483 | |
079cdb61 | 484 | static void spu_prio_wait(struct spu_context *ctx) |
8b3d6663 | 485 | { |
a68cf983 | 486 | DEFINE_WAIT(wait); |
8b3d6663 | 487 | |
4e0f4ed0 | 488 | spin_lock(&spu_prio->runq_lock); |
079cdb61 | 489 | prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE); |
a68cf983 | 490 | if (!signal_pending(current)) { |
4e0f4ed0 LB |
491 | __spu_add_to_rq(ctx); |
492 | spin_unlock(&spu_prio->runq_lock); | |
650f8b02 | 493 | mutex_unlock(&ctx->state_mutex); |
a68cf983 | 494 | schedule(); |
650f8b02 | 495 | mutex_lock(&ctx->state_mutex); |
4e0f4ed0 LB |
496 | spin_lock(&spu_prio->runq_lock); |
497 | __spu_del_from_rq(ctx); | |
8b3d6663 | 498 | } |
4e0f4ed0 | 499 | spin_unlock(&spu_prio->runq_lock); |
079cdb61 CH |
500 | __set_current_state(TASK_RUNNING); |
501 | remove_wait_queue(&ctx->stop_wq, &wait); | |
8b3d6663 AB |
502 | } |
503 | ||
079cdb61 | 504 | static struct spu *spu_get_idle(struct spu_context *ctx) |
a68cf983 MN |
505 | { |
506 | struct spu *spu = NULL; | |
507 | int node = cpu_to_node(raw_smp_processor_id()); | |
508 | int n; | |
509 | ||
cbc23d3e AB |
510 | spu = affinity_check(ctx); |
511 | if (spu) | |
512 | return spu_alloc_spu(spu); | |
513 | ||
a68cf983 MN |
514 | for (n = 0; n < MAX_NUMNODES; n++, node++) { |
515 | node = (node < MAX_NUMNODES) ? node : 0; | |
ea1ae594 | 516 | if (!node_allowed(ctx, node)) |
a68cf983 MN |
517 | continue; |
518 | spu = spu_alloc_node(node); | |
519 | if (spu) | |
520 | break; | |
521 | } | |
522 | return spu; | |
523 | } | |
8b3d6663 | 524 | |
52f04fcf CH |
525 | /** |
526 | * find_victim - find a lower priority context to preempt | |
527 | * @ctx: canidate context for running | |
528 | * | |
529 | * Returns the freed physical spu to run the new context on. | |
530 | */ | |
531 | static struct spu *find_victim(struct spu_context *ctx) | |
532 | { | |
533 | struct spu_context *victim = NULL; | |
534 | struct spu *spu; | |
535 | int node, n; | |
536 | ||
537 | /* | |
538 | * Look for a possible preemption candidate on the local node first. | |
539 | * If there is no candidate look at the other nodes. This isn't | |
540 | * exactly fair, but so far the whole spu schedule tries to keep | |
541 | * a strong node affinity. We might want to fine-tune this in | |
542 | * the future. | |
543 | */ | |
544 | restart: | |
545 | node = cpu_to_node(raw_smp_processor_id()); | |
546 | for (n = 0; n < MAX_NUMNODES; n++, node++) { | |
547 | node = (node < MAX_NUMNODES) ? node : 0; | |
ea1ae594 | 548 | if (!node_allowed(ctx, node)) |
52f04fcf CH |
549 | continue; |
550 | ||
551 | mutex_lock(&spu_prio->active_mutex[node]); | |
552 | list_for_each_entry(spu, &spu_prio->active_list[node], list) { | |
553 | struct spu_context *tmp = spu->ctx; | |
554 | ||
fe443ef2 CH |
555 | if (tmp->prio > ctx->prio && |
556 | (!victim || tmp->prio > victim->prio)) | |
52f04fcf CH |
557 | victim = spu->ctx; |
558 | } | |
559 | mutex_unlock(&spu_prio->active_mutex[node]); | |
560 | ||
561 | if (victim) { | |
562 | /* | |
563 | * This nests ctx->state_mutex, but we always lock | |
564 | * higher priority contexts before lower priority | |
565 | * ones, so this is safe until we introduce | |
566 | * priority inheritance schemes. | |
567 | */ | |
568 | if (!mutex_trylock(&victim->state_mutex)) { | |
569 | victim = NULL; | |
570 | goto restart; | |
571 | } | |
572 | ||
573 | spu = victim->spu; | |
574 | if (!spu) { | |
575 | /* | |
576 | * This race can happen because we've dropped | |
577 | * the active list mutex. No a problem, just | |
578 | * restart the search. | |
579 | */ | |
580 | mutex_unlock(&victim->state_mutex); | |
581 | victim = NULL; | |
582 | goto restart; | |
583 | } | |
37901802 | 584 | spu_remove_from_active_list(spu); |
52f04fcf | 585 | spu_unbind_context(spu, victim); |
e9f8a0b6 | 586 | victim->stats.invol_ctx_switch++; |
fe2f896d | 587 | spu->stats.invol_ctx_switch++; |
52f04fcf | 588 | mutex_unlock(&victim->state_mutex); |
e097b513 CH |
589 | /* |
590 | * We need to break out of the wait loop in spu_run | |
591 | * manually to ensure this context gets put on the | |
592 | * runqueue again ASAP. | |
593 | */ | |
594 | wake_up(&victim->stop_wq); | |
52f04fcf CH |
595 | return spu; |
596 | } | |
597 | } | |
598 | ||
599 | return NULL; | |
600 | } | |
601 | ||
079cdb61 CH |
602 | /** |
603 | * spu_activate - find a free spu for a context and execute it | |
604 | * @ctx: spu context to schedule | |
605 | * @flags: flags (currently ignored) | |
606 | * | |
08873095 | 607 | * Tries to find a free spu to run @ctx. If no free spu is available |
079cdb61 CH |
608 | * add the context to the runqueue so it gets woken up once an spu |
609 | * is available. | |
610 | */ | |
26bec673 | 611 | int spu_activate(struct spu_context *ctx, unsigned long flags) |
8b3d6663 | 612 | { |
079cdb61 CH |
613 | do { |
614 | struct spu *spu; | |
615 | ||
27449971 CH |
616 | /* |
617 | * If there are multiple threads waiting for a single context | |
618 | * only one actually binds the context while the others will | |
619 | * only be able to acquire the state_mutex once the context | |
620 | * already is in runnable state. | |
621 | */ | |
622 | if (ctx->spu) | |
623 | return 0; | |
624 | ||
079cdb61 | 625 | spu = spu_get_idle(ctx); |
52f04fcf CH |
626 | /* |
627 | * If this is a realtime thread we try to get it running by | |
628 | * preempting a lower priority thread. | |
629 | */ | |
fe443ef2 | 630 | if (!spu && rt_prio(ctx->prio)) |
52f04fcf | 631 | spu = find_victim(ctx); |
079cdb61 | 632 | if (spu) { |
202557d2 | 633 | spu_bind_context(spu, ctx); |
37901802 | 634 | spu_add_to_active_list(spu); |
079cdb61 | 635 | return 0; |
a68cf983 | 636 | } |
079cdb61 | 637 | |
50b520d4 | 638 | spu_prio_wait(ctx); |
079cdb61 CH |
639 | } while (!signal_pending(current)); |
640 | ||
641 | return -ERESTARTSYS; | |
8b3d6663 AB |
642 | } |
643 | ||
bb5db29a CH |
644 | /** |
645 | * grab_runnable_context - try to find a runnable context | |
646 | * | |
647 | * Remove the highest priority context on the runqueue and return it | |
648 | * to the caller. Returns %NULL if no runnable context was found. | |
649 | */ | |
ea1ae594 | 650 | static struct spu_context *grab_runnable_context(int prio, int node) |
bb5db29a | 651 | { |
ea1ae594 | 652 | struct spu_context *ctx; |
bb5db29a CH |
653 | int best; |
654 | ||
655 | spin_lock(&spu_prio->runq_lock); | |
7e90b749 | 656 | best = find_first_bit(spu_prio->bitmap, prio); |
ea1ae594 | 657 | while (best < prio) { |
bb5db29a CH |
658 | struct list_head *rq = &spu_prio->runq[best]; |
659 | ||
ea1ae594 CH |
660 | list_for_each_entry(ctx, rq, rq) { |
661 | /* XXX(hch): check for affinity here aswell */ | |
662 | if (__node_allowed(ctx, node)) { | |
663 | __spu_del_from_rq(ctx); | |
664 | goto found; | |
665 | } | |
666 | } | |
667 | best++; | |
bb5db29a | 668 | } |
ea1ae594 CH |
669 | ctx = NULL; |
670 | found: | |
bb5db29a | 671 | spin_unlock(&spu_prio->runq_lock); |
bb5db29a CH |
672 | return ctx; |
673 | } | |
674 | ||
675 | static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio) | |
676 | { | |
677 | struct spu *spu = ctx->spu; | |
678 | struct spu_context *new = NULL; | |
679 | ||
680 | if (spu) { | |
ea1ae594 | 681 | new = grab_runnable_context(max_prio, spu->node); |
bb5db29a | 682 | if (new || force) { |
37901802 | 683 | spu_remove_from_active_list(spu); |
bb5db29a | 684 | spu_unbind_context(spu, ctx); |
e9f8a0b6 | 685 | ctx->stats.vol_ctx_switch++; |
fe2f896d | 686 | spu->stats.vol_ctx_switch++; |
bb5db29a CH |
687 | spu_free(spu); |
688 | if (new) | |
689 | wake_up(&new->stop_wq); | |
690 | } | |
691 | ||
692 | } | |
693 | ||
694 | return new != NULL; | |
695 | } | |
696 | ||
678b2ff1 CH |
697 | /** |
698 | * spu_deactivate - unbind a context from it's physical spu | |
699 | * @ctx: spu context to unbind | |
700 | * | |
701 | * Unbind @ctx from the physical spu it is running on and schedule | |
702 | * the highest priority context to run on the freed physical spu. | |
703 | */ | |
8b3d6663 AB |
704 | void spu_deactivate(struct spu_context *ctx) |
705 | { | |
bb5db29a | 706 | __spu_deactivate(ctx, 1, MAX_PRIO); |
8b3d6663 AB |
707 | } |
708 | ||
ae7b4c52 CH |
709 | /** |
710 | * spu_yield - yield a physical spu if others are waiting | |
711 | * @ctx: spu context to yield | |
712 | * | |
713 | * Check if there is a higher priority context waiting and if yes | |
714 | * unbind @ctx from the physical spu and schedule the highest | |
715 | * priority context to run on the freed physical spu instead. | |
716 | */ | |
8b3d6663 AB |
717 | void spu_yield(struct spu_context *ctx) |
718 | { | |
e5c0b9ec CH |
719 | if (!(ctx->flags & SPU_CREATE_NOSCHED)) { |
720 | mutex_lock(&ctx->state_mutex); | |
27ec41d3 | 721 | __spu_deactivate(ctx, 0, MAX_PRIO); |
e5c0b9ec CH |
722 | mutex_unlock(&ctx->state_mutex); |
723 | } | |
bb5db29a | 724 | } |
8b3d6663 | 725 | |
37901802 | 726 | static void spusched_tick(struct spu_context *ctx) |
bb5db29a | 727 | { |
df09cf3e CH |
728 | if (ctx->flags & SPU_CREATE_NOSCHED) |
729 | return; | |
730 | if (ctx->policy == SCHED_FIFO) | |
731 | return; | |
732 | ||
733 | if (--ctx->time_slice) | |
37901802 | 734 | return; |
bb5db29a CH |
735 | |
736 | /* | |
37901802 CH |
737 | * Unfortunately active_mutex ranks outside of state_mutex, so |
738 | * we have to trylock here. If we fail give the context another | |
739 | * tick and try again. | |
bb5db29a | 740 | */ |
37901802 | 741 | if (mutex_trylock(&ctx->state_mutex)) { |
7022543e | 742 | struct spu *spu = ctx->spu; |
ea1ae594 CH |
743 | struct spu_context *new; |
744 | ||
745 | new = grab_runnable_context(ctx->prio + 1, spu->node); | |
37901802 | 746 | if (new) { |
bb5db29a | 747 | |
37901802 CH |
748 | __spu_remove_from_active_list(spu); |
749 | spu_unbind_context(spu, ctx); | |
e9f8a0b6 | 750 | ctx->stats.invol_ctx_switch++; |
fe2f896d | 751 | spu->stats.invol_ctx_switch++; |
37901802 CH |
752 | spu_free(spu); |
753 | wake_up(&new->stop_wq); | |
754 | /* | |
755 | * We need to break out of the wait loop in | |
756 | * spu_run manually to ensure this context | |
757 | * gets put on the runqueue again ASAP. | |
758 | */ | |
759 | wake_up(&ctx->stop_wq); | |
760 | } | |
fe443ef2 | 761 | spu_set_timeslice(ctx); |
37901802 | 762 | mutex_unlock(&ctx->state_mutex); |
bb5db29a | 763 | } else { |
37901802 | 764 | ctx->time_slice++; |
8b3d6663 | 765 | } |
8b3d6663 AB |
766 | } |
767 | ||
65de66f0 CH |
768 | /** |
769 | * count_active_contexts - count nr of active tasks | |
770 | * | |
771 | * Return the number of tasks currently running or waiting to run. | |
772 | * | |
773 | * Note that we don't take runq_lock / active_mutex here. Reading | |
774 | * a single 32bit value is atomic on powerpc, and we don't care | |
775 | * about memory ordering issues here. | |
776 | */ | |
777 | static unsigned long count_active_contexts(void) | |
778 | { | |
779 | int nr_active = 0, node; | |
780 | ||
781 | for (node = 0; node < MAX_NUMNODES; node++) | |
782 | nr_active += spu_prio->nr_active[node]; | |
783 | nr_active += spu_prio->nr_waiting; | |
784 | ||
785 | return nr_active; | |
786 | } | |
787 | ||
788 | /** | |
789 | * spu_calc_load - given tick count, update the avenrun load estimates. | |
790 | * @tick: tick count | |
791 | * | |
792 | * No locking against reading these values from userspace, as for | |
793 | * the CPU loadavg code. | |
794 | */ | |
795 | static void spu_calc_load(unsigned long ticks) | |
796 | { | |
797 | unsigned long active_tasks; /* fixed-point */ | |
798 | static int count = LOAD_FREQ; | |
799 | ||
800 | count -= ticks; | |
801 | ||
802 | if (unlikely(count < 0)) { | |
803 | active_tasks = count_active_contexts() * FIXED_1; | |
804 | do { | |
805 | CALC_LOAD(spu_avenrun[0], EXP_1, active_tasks); | |
806 | CALC_LOAD(spu_avenrun[1], EXP_5, active_tasks); | |
807 | CALC_LOAD(spu_avenrun[2], EXP_15, active_tasks); | |
808 | count += LOAD_FREQ; | |
809 | } while (count < 0); | |
810 | } | |
811 | } | |
812 | ||
37901802 CH |
813 | static void spusched_wake(unsigned long data) |
814 | { | |
815 | mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK); | |
816 | wake_up_process(spusched_task); | |
65de66f0 | 817 | spu_calc_load(SPUSCHED_TICK); |
37901802 CH |
818 | } |
819 | ||
820 | static int spusched_thread(void *unused) | |
821 | { | |
822 | struct spu *spu, *next; | |
823 | int node; | |
824 | ||
37901802 CH |
825 | while (!kthread_should_stop()) { |
826 | set_current_state(TASK_INTERRUPTIBLE); | |
827 | schedule(); | |
828 | for (node = 0; node < MAX_NUMNODES; node++) { | |
829 | mutex_lock(&spu_prio->active_mutex[node]); | |
830 | list_for_each_entry_safe(spu, next, | |
831 | &spu_prio->active_list[node], | |
832 | list) | |
833 | spusched_tick(spu->ctx); | |
834 | mutex_unlock(&spu_prio->active_mutex[node]); | |
835 | } | |
836 | } | |
837 | ||
37901802 CH |
838 | return 0; |
839 | } | |
840 | ||
65de66f0 CH |
841 | #define LOAD_INT(x) ((x) >> FSHIFT) |
842 | #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100) | |
843 | ||
844 | static int show_spu_loadavg(struct seq_file *s, void *private) | |
845 | { | |
846 | int a, b, c; | |
847 | ||
848 | a = spu_avenrun[0] + (FIXED_1/200); | |
849 | b = spu_avenrun[1] + (FIXED_1/200); | |
850 | c = spu_avenrun[2] + (FIXED_1/200); | |
851 | ||
852 | /* | |
853 | * Note that last_pid doesn't really make much sense for the | |
854 | * SPU loadavg (it even seems very odd on the CPU side..), | |
855 | * but we include it here to have a 100% compatible interface. | |
856 | */ | |
857 | seq_printf(s, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n", | |
858 | LOAD_INT(a), LOAD_FRAC(a), | |
859 | LOAD_INT(b), LOAD_FRAC(b), | |
860 | LOAD_INT(c), LOAD_FRAC(c), | |
861 | count_active_contexts(), | |
862 | atomic_read(&nr_spu_contexts), | |
863 | current->nsproxy->pid_ns->last_pid); | |
864 | return 0; | |
865 | } | |
866 | ||
867 | static int spu_loadavg_open(struct inode *inode, struct file *file) | |
868 | { | |
869 | return single_open(file, show_spu_loadavg, NULL); | |
870 | } | |
871 | ||
872 | static const struct file_operations spu_loadavg_fops = { | |
873 | .open = spu_loadavg_open, | |
874 | .read = seq_read, | |
875 | .llseek = seq_lseek, | |
876 | .release = single_release, | |
877 | }; | |
878 | ||
8b3d6663 AB |
879 | int __init spu_sched_init(void) |
880 | { | |
65de66f0 CH |
881 | struct proc_dir_entry *entry; |
882 | int err = -ENOMEM, i; | |
8b3d6663 | 883 | |
a68cf983 | 884 | spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL); |
37901802 | 885 | if (!spu_prio) |
65de66f0 | 886 | goto out; |
37901802 | 887 | |
8b3d6663 | 888 | for (i = 0; i < MAX_PRIO; i++) { |
079cdb61 | 889 | INIT_LIST_HEAD(&spu_prio->runq[i]); |
a68cf983 | 890 | __clear_bit(i, spu_prio->bitmap); |
8b3d6663 | 891 | } |
a68cf983 MN |
892 | for (i = 0; i < MAX_NUMNODES; i++) { |
893 | mutex_init(&spu_prio->active_mutex[i]); | |
894 | INIT_LIST_HEAD(&spu_prio->active_list[i]); | |
8b3d6663 | 895 | } |
079cdb61 | 896 | spin_lock_init(&spu_prio->runq_lock); |
37901802 | 897 | |
c77239b8 CH |
898 | setup_timer(&spusched_timer, spusched_wake, 0); |
899 | ||
37901802 CH |
900 | spusched_task = kthread_run(spusched_thread, NULL, "spusched"); |
901 | if (IS_ERR(spusched_task)) { | |
65de66f0 CH |
902 | err = PTR_ERR(spusched_task); |
903 | goto out_free_spu_prio; | |
37901802 | 904 | } |
f3f59bec | 905 | |
65de66f0 CH |
906 | entry = create_proc_entry("spu_loadavg", 0, NULL); |
907 | if (!entry) | |
908 | goto out_stop_kthread; | |
909 | entry->proc_fops = &spu_loadavg_fops; | |
910 | ||
f3f59bec JK |
911 | pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n", |
912 | SPUSCHED_TICK, MIN_SPU_TIMESLICE, DEF_SPU_TIMESLICE); | |
8b3d6663 | 913 | return 0; |
37901802 | 914 | |
65de66f0 CH |
915 | out_stop_kthread: |
916 | kthread_stop(spusched_task); | |
917 | out_free_spu_prio: | |
918 | kfree(spu_prio); | |
919 | out: | |
920 | return err; | |
8b3d6663 AB |
921 | } |
922 | ||
d1450317 | 923 | void spu_sched_exit(void) |
8b3d6663 | 924 | { |
a68cf983 MN |
925 | struct spu *spu, *tmp; |
926 | int node; | |
927 | ||
65de66f0 CH |
928 | remove_proc_entry("spu_loadavg", NULL); |
929 | ||
c77239b8 | 930 | del_timer_sync(&spusched_timer); |
37901802 CH |
931 | kthread_stop(spusched_task); |
932 | ||
a68cf983 MN |
933 | for (node = 0; node < MAX_NUMNODES; node++) { |
934 | mutex_lock(&spu_prio->active_mutex[node]); | |
935 | list_for_each_entry_safe(spu, tmp, &spu_prio->active_list[node], | |
936 | list) { | |
937 | list_del_init(&spu->list); | |
938 | spu_free(spu); | |
939 | } | |
940 | mutex_unlock(&spu_prio->active_mutex[node]); | |
8b3d6663 | 941 | } |
a68cf983 | 942 | kfree(spu_prio); |
8b3d6663 | 943 | } |