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 CH |
233 | |
234 | ctx->stats.slb_flt_base = spu->stats.slb_flt; | |
235 | ctx->stats.class2_intr_base = spu->stats.class2_intr; | |
236 | ||
8b3d6663 AB |
237 | spu->ctx = ctx; |
238 | spu->flags = 0; | |
239 | ctx->spu = spu; | |
240 | ctx->ops = &spu_hw_ops; | |
241 | spu->pid = current->pid; | |
94b2a439 | 242 | spu_associate_mm(spu, ctx->owner); |
8b3d6663 AB |
243 | spu->ibox_callback = spufs_ibox_callback; |
244 | spu->wbox_callback = spufs_wbox_callback; | |
5110459f | 245 | spu->stop_callback = spufs_stop_callback; |
a33a7d73 | 246 | spu->mfc_callback = spufs_mfc_callback; |
9add11da | 247 | spu->dma_callback = spufs_dma_callback; |
8b3d6663 | 248 | mb(); |
5110459f | 249 | spu_unmap_mappings(ctx); |
8b3d6663 | 250 | spu_restore(&ctx->csa, spu); |
2a911f0b | 251 | spu->timestamp = jiffies; |
a68cf983 | 252 | spu_cpu_affinity_set(spu, raw_smp_processor_id()); |
86767277 | 253 | spu_switch_notify(spu, ctx); |
81998baf | 254 | ctx->state = SPU_STATE_RUNNABLE; |
27ec41d3 AD |
255 | |
256 | spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED); | |
8b3d6663 AB |
257 | } |
258 | ||
202557d2 CH |
259 | /** |
260 | * spu_unbind_context - unbind spu context from physical spu | |
261 | * @spu: physical spu to unbind from | |
262 | * @ctx: context to unbind | |
202557d2 | 263 | */ |
678b2ff1 | 264 | static void spu_unbind_context(struct spu *spu, struct spu_context *ctx) |
8b3d6663 | 265 | { |
a68cf983 MN |
266 | pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__, |
267 | spu->pid, spu->number, spu->node); | |
27ec41d3 | 268 | spuctx_switch_state(ctx, SPU_UTIL_SYSTEM); |
fe2f896d | 269 | |
86767277 | 270 | spu_switch_notify(spu, NULL); |
5110459f | 271 | spu_unmap_mappings(ctx); |
8b3d6663 | 272 | spu_save(&ctx->csa, spu); |
2a911f0b | 273 | spu->timestamp = jiffies; |
8b3d6663 AB |
274 | ctx->state = SPU_STATE_SAVED; |
275 | spu->ibox_callback = NULL; | |
276 | spu->wbox_callback = NULL; | |
5110459f | 277 | spu->stop_callback = NULL; |
a33a7d73 | 278 | spu->mfc_callback = NULL; |
9add11da | 279 | spu->dma_callback = NULL; |
94b2a439 | 280 | spu_associate_mm(spu, NULL); |
8b3d6663 | 281 | spu->pid = 0; |
8b3d6663 | 282 | ctx->ops = &spu_backing_ops; |
2a911f0b | 283 | spu->flags = 0; |
8b3d6663 | 284 | spu->ctx = NULL; |
e9f8a0b6 CH |
285 | |
286 | ctx->stats.slb_flt += | |
287 | (spu->stats.slb_flt - ctx->stats.slb_flt_base); | |
288 | ctx->stats.class2_intr += | |
289 | (spu->stats.class2_intr - ctx->stats.class2_intr_base); | |
27ec41d3 AD |
290 | |
291 | /* This maps the underlying spu state to idle */ | |
292 | spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED); | |
293 | ctx->spu = NULL; | |
8b3d6663 AB |
294 | } |
295 | ||
079cdb61 CH |
296 | /** |
297 | * spu_add_to_rq - add a context to the runqueue | |
298 | * @ctx: context to add | |
299 | */ | |
4e0f4ed0 | 300 | static void __spu_add_to_rq(struct spu_context *ctx) |
8b3d6663 | 301 | { |
27449971 CH |
302 | /* |
303 | * Unfortunately this code path can be called from multiple threads | |
304 | * on behalf of a single context due to the way the problem state | |
305 | * mmap support works. | |
306 | * | |
307 | * Fortunately we need to wake up all these threads at the same time | |
308 | * and can simply skip the runqueue addition for every but the first | |
309 | * thread getting into this codepath. | |
310 | * | |
311 | * It's still quite hacky, and long-term we should proxy all other | |
312 | * threads through the owner thread so that spu_run is in control | |
313 | * of all the scheduling activity for a given context. | |
314 | */ | |
315 | if (list_empty(&ctx->rq)) { | |
316 | list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]); | |
317 | set_bit(ctx->prio, spu_prio->bitmap); | |
318 | if (!spu_prio->nr_waiting++) | |
319 | __mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK); | |
320 | } | |
2a911f0b | 321 | } |
5110459f | 322 | |
4e0f4ed0 | 323 | static void __spu_del_from_rq(struct spu_context *ctx) |
a475c2f4 | 324 | { |
4e0f4ed0 LB |
325 | int prio = ctx->prio; |
326 | ||
65de66f0 | 327 | if (!list_empty(&ctx->rq)) { |
c77239b8 CH |
328 | if (!--spu_prio->nr_waiting) |
329 | del_timer(&spusched_timer); | |
a475c2f4 | 330 | list_del_init(&ctx->rq); |
c77239b8 CH |
331 | |
332 | if (list_empty(&spu_prio->runq[prio])) | |
333 | clear_bit(prio, spu_prio->bitmap); | |
65de66f0 | 334 | } |
079cdb61 | 335 | } |
a68cf983 | 336 | |
079cdb61 | 337 | static void spu_prio_wait(struct spu_context *ctx) |
8b3d6663 | 338 | { |
a68cf983 | 339 | DEFINE_WAIT(wait); |
8b3d6663 | 340 | |
4e0f4ed0 | 341 | spin_lock(&spu_prio->runq_lock); |
079cdb61 | 342 | prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE); |
a68cf983 | 343 | if (!signal_pending(current)) { |
4e0f4ed0 LB |
344 | __spu_add_to_rq(ctx); |
345 | spin_unlock(&spu_prio->runq_lock); | |
650f8b02 | 346 | mutex_unlock(&ctx->state_mutex); |
a68cf983 | 347 | schedule(); |
650f8b02 | 348 | mutex_lock(&ctx->state_mutex); |
4e0f4ed0 LB |
349 | spin_lock(&spu_prio->runq_lock); |
350 | __spu_del_from_rq(ctx); | |
8b3d6663 | 351 | } |
4e0f4ed0 | 352 | spin_unlock(&spu_prio->runq_lock); |
079cdb61 CH |
353 | __set_current_state(TASK_RUNNING); |
354 | remove_wait_queue(&ctx->stop_wq, &wait); | |
8b3d6663 AB |
355 | } |
356 | ||
079cdb61 | 357 | static struct spu *spu_get_idle(struct spu_context *ctx) |
a68cf983 MN |
358 | { |
359 | struct spu *spu = NULL; | |
360 | int node = cpu_to_node(raw_smp_processor_id()); | |
361 | int n; | |
362 | ||
363 | for (n = 0; n < MAX_NUMNODES; n++, node++) { | |
364 | node = (node < MAX_NUMNODES) ? node : 0; | |
ea1ae594 | 365 | if (!node_allowed(ctx, node)) |
a68cf983 MN |
366 | continue; |
367 | spu = spu_alloc_node(node); | |
368 | if (spu) | |
369 | break; | |
370 | } | |
371 | return spu; | |
372 | } | |
8b3d6663 | 373 | |
52f04fcf CH |
374 | /** |
375 | * find_victim - find a lower priority context to preempt | |
376 | * @ctx: canidate context for running | |
377 | * | |
378 | * Returns the freed physical spu to run the new context on. | |
379 | */ | |
380 | static struct spu *find_victim(struct spu_context *ctx) | |
381 | { | |
382 | struct spu_context *victim = NULL; | |
383 | struct spu *spu; | |
384 | int node, n; | |
385 | ||
386 | /* | |
387 | * Look for a possible preemption candidate on the local node first. | |
388 | * If there is no candidate look at the other nodes. This isn't | |
389 | * exactly fair, but so far the whole spu schedule tries to keep | |
390 | * a strong node affinity. We might want to fine-tune this in | |
391 | * the future. | |
392 | */ | |
393 | restart: | |
394 | node = cpu_to_node(raw_smp_processor_id()); | |
395 | for (n = 0; n < MAX_NUMNODES; n++, node++) { | |
396 | node = (node < MAX_NUMNODES) ? node : 0; | |
ea1ae594 | 397 | if (!node_allowed(ctx, node)) |
52f04fcf CH |
398 | continue; |
399 | ||
400 | mutex_lock(&spu_prio->active_mutex[node]); | |
401 | list_for_each_entry(spu, &spu_prio->active_list[node], list) { | |
402 | struct spu_context *tmp = spu->ctx; | |
403 | ||
fe443ef2 CH |
404 | if (tmp->prio > ctx->prio && |
405 | (!victim || tmp->prio > victim->prio)) | |
52f04fcf CH |
406 | victim = spu->ctx; |
407 | } | |
408 | mutex_unlock(&spu_prio->active_mutex[node]); | |
409 | ||
410 | if (victim) { | |
411 | /* | |
412 | * This nests ctx->state_mutex, but we always lock | |
413 | * higher priority contexts before lower priority | |
414 | * ones, so this is safe until we introduce | |
415 | * priority inheritance schemes. | |
416 | */ | |
417 | if (!mutex_trylock(&victim->state_mutex)) { | |
418 | victim = NULL; | |
419 | goto restart; | |
420 | } | |
421 | ||
422 | spu = victim->spu; | |
423 | if (!spu) { | |
424 | /* | |
425 | * This race can happen because we've dropped | |
426 | * the active list mutex. No a problem, just | |
427 | * restart the search. | |
428 | */ | |
429 | mutex_unlock(&victim->state_mutex); | |
430 | victim = NULL; | |
431 | goto restart; | |
432 | } | |
37901802 | 433 | spu_remove_from_active_list(spu); |
52f04fcf | 434 | spu_unbind_context(spu, victim); |
e9f8a0b6 | 435 | victim->stats.invol_ctx_switch++; |
fe2f896d | 436 | spu->stats.invol_ctx_switch++; |
52f04fcf | 437 | mutex_unlock(&victim->state_mutex); |
e097b513 CH |
438 | /* |
439 | * We need to break out of the wait loop in spu_run | |
440 | * manually to ensure this context gets put on the | |
441 | * runqueue again ASAP. | |
442 | */ | |
443 | wake_up(&victim->stop_wq); | |
52f04fcf CH |
444 | return spu; |
445 | } | |
446 | } | |
447 | ||
448 | return NULL; | |
449 | } | |
450 | ||
079cdb61 CH |
451 | /** |
452 | * spu_activate - find a free spu for a context and execute it | |
453 | * @ctx: spu context to schedule | |
454 | * @flags: flags (currently ignored) | |
455 | * | |
08873095 | 456 | * Tries to find a free spu to run @ctx. If no free spu is available |
079cdb61 CH |
457 | * add the context to the runqueue so it gets woken up once an spu |
458 | * is available. | |
459 | */ | |
26bec673 | 460 | int spu_activate(struct spu_context *ctx, unsigned long flags) |
8b3d6663 | 461 | { |
079cdb61 CH |
462 | do { |
463 | struct spu *spu; | |
464 | ||
27449971 CH |
465 | /* |
466 | * If there are multiple threads waiting for a single context | |
467 | * only one actually binds the context while the others will | |
468 | * only be able to acquire the state_mutex once the context | |
469 | * already is in runnable state. | |
470 | */ | |
471 | if (ctx->spu) | |
472 | return 0; | |
473 | ||
079cdb61 | 474 | spu = spu_get_idle(ctx); |
52f04fcf CH |
475 | /* |
476 | * If this is a realtime thread we try to get it running by | |
477 | * preempting a lower priority thread. | |
478 | */ | |
fe443ef2 | 479 | if (!spu && rt_prio(ctx->prio)) |
52f04fcf | 480 | spu = find_victim(ctx); |
079cdb61 | 481 | if (spu) { |
202557d2 | 482 | spu_bind_context(spu, ctx); |
37901802 | 483 | spu_add_to_active_list(spu); |
079cdb61 | 484 | return 0; |
a68cf983 | 485 | } |
079cdb61 | 486 | |
50b520d4 | 487 | spu_prio_wait(ctx); |
079cdb61 CH |
488 | } while (!signal_pending(current)); |
489 | ||
490 | return -ERESTARTSYS; | |
8b3d6663 AB |
491 | } |
492 | ||
bb5db29a CH |
493 | /** |
494 | * grab_runnable_context - try to find a runnable context | |
495 | * | |
496 | * Remove the highest priority context on the runqueue and return it | |
497 | * to the caller. Returns %NULL if no runnable context was found. | |
498 | */ | |
ea1ae594 | 499 | static struct spu_context *grab_runnable_context(int prio, int node) |
bb5db29a | 500 | { |
ea1ae594 | 501 | struct spu_context *ctx; |
bb5db29a CH |
502 | int best; |
503 | ||
504 | spin_lock(&spu_prio->runq_lock); | |
505 | best = sched_find_first_bit(spu_prio->bitmap); | |
ea1ae594 | 506 | while (best < prio) { |
bb5db29a CH |
507 | struct list_head *rq = &spu_prio->runq[best]; |
508 | ||
ea1ae594 CH |
509 | list_for_each_entry(ctx, rq, rq) { |
510 | /* XXX(hch): check for affinity here aswell */ | |
511 | if (__node_allowed(ctx, node)) { | |
512 | __spu_del_from_rq(ctx); | |
513 | goto found; | |
514 | } | |
515 | } | |
516 | best++; | |
bb5db29a | 517 | } |
ea1ae594 CH |
518 | ctx = NULL; |
519 | found: | |
bb5db29a | 520 | spin_unlock(&spu_prio->runq_lock); |
bb5db29a CH |
521 | return ctx; |
522 | } | |
523 | ||
524 | static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio) | |
525 | { | |
526 | struct spu *spu = ctx->spu; | |
527 | struct spu_context *new = NULL; | |
528 | ||
529 | if (spu) { | |
ea1ae594 | 530 | new = grab_runnable_context(max_prio, spu->node); |
bb5db29a | 531 | if (new || force) { |
37901802 | 532 | spu_remove_from_active_list(spu); |
bb5db29a | 533 | spu_unbind_context(spu, ctx); |
e9f8a0b6 | 534 | ctx->stats.vol_ctx_switch++; |
fe2f896d | 535 | spu->stats.vol_ctx_switch++; |
bb5db29a CH |
536 | spu_free(spu); |
537 | if (new) | |
538 | wake_up(&new->stop_wq); | |
539 | } | |
540 | ||
541 | } | |
542 | ||
543 | return new != NULL; | |
544 | } | |
545 | ||
678b2ff1 CH |
546 | /** |
547 | * spu_deactivate - unbind a context from it's physical spu | |
548 | * @ctx: spu context to unbind | |
549 | * | |
550 | * Unbind @ctx from the physical spu it is running on and schedule | |
551 | * the highest priority context to run on the freed physical spu. | |
552 | */ | |
8b3d6663 AB |
553 | void spu_deactivate(struct spu_context *ctx) |
554 | { | |
bb5db29a | 555 | __spu_deactivate(ctx, 1, MAX_PRIO); |
8b3d6663 AB |
556 | } |
557 | ||
ae7b4c52 CH |
558 | /** |
559 | * spu_yield - yield a physical spu if others are waiting | |
560 | * @ctx: spu context to yield | |
561 | * | |
562 | * Check if there is a higher priority context waiting and if yes | |
563 | * unbind @ctx from the physical spu and schedule the highest | |
564 | * priority context to run on the freed physical spu instead. | |
565 | */ | |
8b3d6663 AB |
566 | void spu_yield(struct spu_context *ctx) |
567 | { | |
e5c0b9ec CH |
568 | if (!(ctx->flags & SPU_CREATE_NOSCHED)) { |
569 | mutex_lock(&ctx->state_mutex); | |
27ec41d3 | 570 | __spu_deactivate(ctx, 0, MAX_PRIO); |
e5c0b9ec CH |
571 | mutex_unlock(&ctx->state_mutex); |
572 | } | |
bb5db29a | 573 | } |
8b3d6663 | 574 | |
37901802 | 575 | static void spusched_tick(struct spu_context *ctx) |
bb5db29a | 576 | { |
df09cf3e CH |
577 | if (ctx->flags & SPU_CREATE_NOSCHED) |
578 | return; | |
579 | if (ctx->policy == SCHED_FIFO) | |
580 | return; | |
581 | ||
582 | if (--ctx->time_slice) | |
37901802 | 583 | return; |
bb5db29a CH |
584 | |
585 | /* | |
37901802 CH |
586 | * Unfortunately active_mutex ranks outside of state_mutex, so |
587 | * we have to trylock here. If we fail give the context another | |
588 | * tick and try again. | |
bb5db29a | 589 | */ |
37901802 | 590 | if (mutex_trylock(&ctx->state_mutex)) { |
7022543e | 591 | struct spu *spu = ctx->spu; |
ea1ae594 CH |
592 | struct spu_context *new; |
593 | ||
594 | new = grab_runnable_context(ctx->prio + 1, spu->node); | |
37901802 | 595 | if (new) { |
bb5db29a | 596 | |
37901802 CH |
597 | __spu_remove_from_active_list(spu); |
598 | spu_unbind_context(spu, ctx); | |
e9f8a0b6 | 599 | ctx->stats.invol_ctx_switch++; |
fe2f896d | 600 | spu->stats.invol_ctx_switch++; |
37901802 CH |
601 | spu_free(spu); |
602 | wake_up(&new->stop_wq); | |
603 | /* | |
604 | * We need to break out of the wait loop in | |
605 | * spu_run manually to ensure this context | |
606 | * gets put on the runqueue again ASAP. | |
607 | */ | |
608 | wake_up(&ctx->stop_wq); | |
609 | } | |
fe443ef2 | 610 | spu_set_timeslice(ctx); |
37901802 | 611 | mutex_unlock(&ctx->state_mutex); |
bb5db29a | 612 | } else { |
37901802 | 613 | ctx->time_slice++; |
8b3d6663 | 614 | } |
8b3d6663 AB |
615 | } |
616 | ||
65de66f0 CH |
617 | /** |
618 | * count_active_contexts - count nr of active tasks | |
619 | * | |
620 | * Return the number of tasks currently running or waiting to run. | |
621 | * | |
622 | * Note that we don't take runq_lock / active_mutex here. Reading | |
623 | * a single 32bit value is atomic on powerpc, and we don't care | |
624 | * about memory ordering issues here. | |
625 | */ | |
626 | static unsigned long count_active_contexts(void) | |
627 | { | |
628 | int nr_active = 0, node; | |
629 | ||
630 | for (node = 0; node < MAX_NUMNODES; node++) | |
631 | nr_active += spu_prio->nr_active[node]; | |
632 | nr_active += spu_prio->nr_waiting; | |
633 | ||
634 | return nr_active; | |
635 | } | |
636 | ||
637 | /** | |
638 | * spu_calc_load - given tick count, update the avenrun load estimates. | |
639 | * @tick: tick count | |
640 | * | |
641 | * No locking against reading these values from userspace, as for | |
642 | * the CPU loadavg code. | |
643 | */ | |
644 | static void spu_calc_load(unsigned long ticks) | |
645 | { | |
646 | unsigned long active_tasks; /* fixed-point */ | |
647 | static int count = LOAD_FREQ; | |
648 | ||
649 | count -= ticks; | |
650 | ||
651 | if (unlikely(count < 0)) { | |
652 | active_tasks = count_active_contexts() * FIXED_1; | |
653 | do { | |
654 | CALC_LOAD(spu_avenrun[0], EXP_1, active_tasks); | |
655 | CALC_LOAD(spu_avenrun[1], EXP_5, active_tasks); | |
656 | CALC_LOAD(spu_avenrun[2], EXP_15, active_tasks); | |
657 | count += LOAD_FREQ; | |
658 | } while (count < 0); | |
659 | } | |
660 | } | |
661 | ||
37901802 CH |
662 | static void spusched_wake(unsigned long data) |
663 | { | |
664 | mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK); | |
665 | wake_up_process(spusched_task); | |
65de66f0 | 666 | spu_calc_load(SPUSCHED_TICK); |
37901802 CH |
667 | } |
668 | ||
669 | static int spusched_thread(void *unused) | |
670 | { | |
671 | struct spu *spu, *next; | |
672 | int node; | |
673 | ||
37901802 CH |
674 | while (!kthread_should_stop()) { |
675 | set_current_state(TASK_INTERRUPTIBLE); | |
676 | schedule(); | |
677 | for (node = 0; node < MAX_NUMNODES; node++) { | |
678 | mutex_lock(&spu_prio->active_mutex[node]); | |
679 | list_for_each_entry_safe(spu, next, | |
680 | &spu_prio->active_list[node], | |
681 | list) | |
682 | spusched_tick(spu->ctx); | |
683 | mutex_unlock(&spu_prio->active_mutex[node]); | |
684 | } | |
685 | } | |
686 | ||
37901802 CH |
687 | return 0; |
688 | } | |
689 | ||
65de66f0 CH |
690 | #define LOAD_INT(x) ((x) >> FSHIFT) |
691 | #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100) | |
692 | ||
693 | static int show_spu_loadavg(struct seq_file *s, void *private) | |
694 | { | |
695 | int a, b, c; | |
696 | ||
697 | a = spu_avenrun[0] + (FIXED_1/200); | |
698 | b = spu_avenrun[1] + (FIXED_1/200); | |
699 | c = spu_avenrun[2] + (FIXED_1/200); | |
700 | ||
701 | /* | |
702 | * Note that last_pid doesn't really make much sense for the | |
703 | * SPU loadavg (it even seems very odd on the CPU side..), | |
704 | * but we include it here to have a 100% compatible interface. | |
705 | */ | |
706 | seq_printf(s, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n", | |
707 | LOAD_INT(a), LOAD_FRAC(a), | |
708 | LOAD_INT(b), LOAD_FRAC(b), | |
709 | LOAD_INT(c), LOAD_FRAC(c), | |
710 | count_active_contexts(), | |
711 | atomic_read(&nr_spu_contexts), | |
712 | current->nsproxy->pid_ns->last_pid); | |
713 | return 0; | |
714 | } | |
715 | ||
716 | static int spu_loadavg_open(struct inode *inode, struct file *file) | |
717 | { | |
718 | return single_open(file, show_spu_loadavg, NULL); | |
719 | } | |
720 | ||
721 | static const struct file_operations spu_loadavg_fops = { | |
722 | .open = spu_loadavg_open, | |
723 | .read = seq_read, | |
724 | .llseek = seq_lseek, | |
725 | .release = single_release, | |
726 | }; | |
727 | ||
8b3d6663 AB |
728 | int __init spu_sched_init(void) |
729 | { | |
65de66f0 CH |
730 | struct proc_dir_entry *entry; |
731 | int err = -ENOMEM, i; | |
8b3d6663 | 732 | |
a68cf983 | 733 | spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL); |
37901802 | 734 | if (!spu_prio) |
65de66f0 | 735 | goto out; |
37901802 | 736 | |
8b3d6663 | 737 | for (i = 0; i < MAX_PRIO; i++) { |
079cdb61 | 738 | INIT_LIST_HEAD(&spu_prio->runq[i]); |
a68cf983 | 739 | __clear_bit(i, spu_prio->bitmap); |
8b3d6663 | 740 | } |
a68cf983 MN |
741 | __set_bit(MAX_PRIO, spu_prio->bitmap); |
742 | for (i = 0; i < MAX_NUMNODES; i++) { | |
743 | mutex_init(&spu_prio->active_mutex[i]); | |
744 | INIT_LIST_HEAD(&spu_prio->active_list[i]); | |
8b3d6663 | 745 | } |
079cdb61 | 746 | spin_lock_init(&spu_prio->runq_lock); |
37901802 | 747 | |
c77239b8 CH |
748 | setup_timer(&spusched_timer, spusched_wake, 0); |
749 | ||
37901802 CH |
750 | spusched_task = kthread_run(spusched_thread, NULL, "spusched"); |
751 | if (IS_ERR(spusched_task)) { | |
65de66f0 CH |
752 | err = PTR_ERR(spusched_task); |
753 | goto out_free_spu_prio; | |
37901802 | 754 | } |
f3f59bec | 755 | |
65de66f0 CH |
756 | entry = create_proc_entry("spu_loadavg", 0, NULL); |
757 | if (!entry) | |
758 | goto out_stop_kthread; | |
759 | entry->proc_fops = &spu_loadavg_fops; | |
760 | ||
f3f59bec JK |
761 | pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n", |
762 | SPUSCHED_TICK, MIN_SPU_TIMESLICE, DEF_SPU_TIMESLICE); | |
8b3d6663 | 763 | return 0; |
37901802 | 764 | |
65de66f0 CH |
765 | out_stop_kthread: |
766 | kthread_stop(spusched_task); | |
767 | out_free_spu_prio: | |
768 | kfree(spu_prio); | |
769 | out: | |
770 | return err; | |
8b3d6663 AB |
771 | } |
772 | ||
d1450317 | 773 | void spu_sched_exit(void) |
8b3d6663 | 774 | { |
a68cf983 MN |
775 | struct spu *spu, *tmp; |
776 | int node; | |
777 | ||
65de66f0 CH |
778 | remove_proc_entry("spu_loadavg", NULL); |
779 | ||
c77239b8 | 780 | del_timer_sync(&spusched_timer); |
37901802 CH |
781 | kthread_stop(spusched_task); |
782 | ||
a68cf983 MN |
783 | for (node = 0; node < MAX_NUMNODES; node++) { |
784 | mutex_lock(&spu_prio->active_mutex[node]); | |
785 | list_for_each_entry_safe(spu, tmp, &spu_prio->active_list[node], | |
786 | list) { | |
787 | list_del_init(&spu->list); | |
788 | spu_free(spu); | |
789 | } | |
790 | mutex_unlock(&spu_prio->active_mutex[node]); | |
8b3d6663 | 791 | } |
a68cf983 | 792 | kfree(spu_prio); |
8b3d6663 | 793 | } |