]>
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; | |
65de66f0 | 54 | int nr_waiting; |
8b3d6663 AB |
55 | }; |
56 | ||
65de66f0 | 57 | static unsigned long spu_avenrun[3]; |
a68cf983 | 58 | static struct spu_prio_array *spu_prio; |
37901802 CH |
59 | static struct task_struct *spusched_task; |
60 | static struct timer_list spusched_timer; | |
8b3d6663 | 61 | |
fe443ef2 CH |
62 | /* |
63 | * Priority of a normal, non-rt, non-niced'd process (aka nice level 0). | |
64 | */ | |
65 | #define NORMAL_PRIO 120 | |
66 | ||
67 | /* | |
68 | * Frequency of the spu scheduler tick. By default we do one SPU scheduler | |
69 | * tick for every 10 CPU scheduler ticks. | |
70 | */ | |
71 | #define SPUSCHED_TICK (10) | |
72 | ||
73 | /* | |
74 | * These are the 'tuning knobs' of the scheduler: | |
75 | * | |
60e24239 JK |
76 | * Minimum timeslice is 5 msecs (or 1 spu scheduler tick, whichever is |
77 | * larger), default timeslice is 100 msecs, maximum timeslice is 800 msecs. | |
fe443ef2 | 78 | */ |
60e24239 JK |
79 | #define MIN_SPU_TIMESLICE max(5 * HZ / (1000 * SPUSCHED_TICK), 1) |
80 | #define DEF_SPU_TIMESLICE (100 * HZ / (1000 * SPUSCHED_TICK)) | |
fe443ef2 CH |
81 | |
82 | #define MAX_USER_PRIO (MAX_PRIO - MAX_RT_PRIO) | |
83 | #define SCALE_PRIO(x, prio) \ | |
84 | max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_SPU_TIMESLICE) | |
85 | ||
86 | /* | |
87 | * scale user-nice values [ -20 ... 0 ... 19 ] to time slice values: | |
88 | * [800ms ... 100ms ... 5ms] | |
89 | * | |
90 | * The higher a thread's priority, the bigger timeslices | |
91 | * it gets during one round of execution. But even the lowest | |
92 | * priority thread gets MIN_TIMESLICE worth of execution time. | |
93 | */ | |
94 | void spu_set_timeslice(struct spu_context *ctx) | |
95 | { | |
96 | if (ctx->prio < NORMAL_PRIO) | |
97 | ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE * 4, ctx->prio); | |
98 | else | |
99 | ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE, ctx->prio); | |
100 | } | |
101 | ||
2cf2b3b4 CH |
102 | /* |
103 | * Update scheduling information from the owning thread. | |
104 | */ | |
105 | void __spu_update_sched_info(struct spu_context *ctx) | |
106 | { | |
476273ad CH |
107 | /* |
108 | * 32-Bit assignment are atomic on powerpc, and we don't care about | |
109 | * memory ordering here because retriving the controlling thread is | |
110 | * per defintion racy. | |
111 | */ | |
112 | ctx->tid = current->pid; | |
113 | ||
2cf2b3b4 CH |
114 | /* |
115 | * We do our own priority calculations, so we normally want | |
116 | * ->static_prio to start with. Unfortunately thies field | |
117 | * contains junk for threads with a realtime scheduling | |
118 | * policy so we have to look at ->prio in this case. | |
119 | */ | |
120 | if (rt_prio(current->prio)) | |
121 | ctx->prio = current->prio; | |
122 | else | |
123 | ctx->prio = current->static_prio; | |
124 | ctx->policy = current->policy; | |
ea1ae594 CH |
125 | |
126 | /* | |
486acd48 | 127 | * A lot of places that don't hold list_mutex poke into |
ea1ae594 CH |
128 | * cpus_allowed, including grab_runnable_context which |
129 | * already holds the runq_lock. So abuse runq_lock | |
130 | * to protect this field aswell. | |
131 | */ | |
132 | spin_lock(&spu_prio->runq_lock); | |
133 | ctx->cpus_allowed = current->cpus_allowed; | |
134 | spin_unlock(&spu_prio->runq_lock); | |
2cf2b3b4 CH |
135 | } |
136 | ||
137 | void spu_update_sched_info(struct spu_context *ctx) | |
138 | { | |
139 | int node = ctx->spu->node; | |
140 | ||
486acd48 | 141 | mutex_lock(&cbe_spu_info[node].list_mutex); |
2cf2b3b4 | 142 | __spu_update_sched_info(ctx); |
486acd48 | 143 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
2cf2b3b4 CH |
144 | } |
145 | ||
ea1ae594 | 146 | static int __node_allowed(struct spu_context *ctx, int node) |
8b3d6663 | 147 | { |
ea1ae594 CH |
148 | if (nr_cpus_node(node)) { |
149 | cpumask_t mask = node_to_cpumask(node); | |
8b3d6663 | 150 | |
ea1ae594 CH |
151 | if (cpus_intersects(mask, ctx->cpus_allowed)) |
152 | return 1; | |
153 | } | |
154 | ||
155 | return 0; | |
156 | } | |
157 | ||
158 | static int node_allowed(struct spu_context *ctx, int node) | |
159 | { | |
160 | int rval; | |
161 | ||
162 | spin_lock(&spu_prio->runq_lock); | |
163 | rval = __node_allowed(ctx, node); | |
164 | spin_unlock(&spu_prio->runq_lock); | |
165 | ||
166 | return rval; | |
8b3d6663 AB |
167 | } |
168 | ||
86767277 AB |
169 | static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier); |
170 | ||
36aaccc1 | 171 | void spu_switch_notify(struct spu *spu, struct spu_context *ctx) |
86767277 AB |
172 | { |
173 | blocking_notifier_call_chain(&spu_switch_notifier, | |
174 | ctx ? ctx->object_id : 0, spu); | |
175 | } | |
176 | ||
36aaccc1 BN |
177 | static void notify_spus_active(void) |
178 | { | |
179 | int node; | |
180 | ||
181 | /* | |
182 | * Wake up the active spu_contexts. | |
183 | * | |
184 | * When the awakened processes see their "notify_active" flag is set, | |
185 | * they will call spu_switch_notify(); | |
186 | */ | |
187 | for_each_online_node(node) { | |
188 | struct spu *spu; | |
486acd48 CH |
189 | |
190 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
191 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { | |
192 | if (spu->alloc_state != SPU_FREE) { | |
193 | struct spu_context *ctx = spu->ctx; | |
194 | set_bit(SPU_SCHED_NOTIFY_ACTIVE, | |
195 | &ctx->sched_flags); | |
196 | mb(); | |
197 | wake_up_all(&ctx->stop_wq); | |
198 | } | |
36aaccc1 | 199 | } |
486acd48 | 200 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
36aaccc1 BN |
201 | } |
202 | } | |
203 | ||
86767277 AB |
204 | int spu_switch_event_register(struct notifier_block * n) |
205 | { | |
36aaccc1 BN |
206 | int ret; |
207 | ret = blocking_notifier_chain_register(&spu_switch_notifier, n); | |
208 | if (!ret) | |
209 | notify_spus_active(); | |
210 | return ret; | |
86767277 | 211 | } |
36aaccc1 | 212 | EXPORT_SYMBOL_GPL(spu_switch_event_register); |
86767277 AB |
213 | |
214 | int spu_switch_event_unregister(struct notifier_block * n) | |
215 | { | |
216 | return blocking_notifier_chain_unregister(&spu_switch_notifier, n); | |
217 | } | |
36aaccc1 | 218 | EXPORT_SYMBOL_GPL(spu_switch_event_unregister); |
86767277 | 219 | |
202557d2 CH |
220 | /** |
221 | * spu_bind_context - bind spu context to physical spu | |
222 | * @spu: physical spu to bind to | |
223 | * @ctx: context to bind | |
224 | */ | |
225 | static void spu_bind_context(struct spu *spu, struct spu_context *ctx) | |
8b3d6663 | 226 | { |
a68cf983 MN |
227 | pr_debug("%s: pid=%d SPU=%d NODE=%d\n", __FUNCTION__, current->pid, |
228 | spu->number, spu->node); | |
27ec41d3 | 229 | spuctx_switch_state(ctx, SPU_UTIL_SYSTEM); |
e9f8a0b6 | 230 | |
aa6d5b20 AB |
231 | if (ctx->flags & SPU_CREATE_NOSCHED) |
232 | atomic_inc(&cbe_spu_info[spu->node].reserved_spus); | |
c5fc8d2a AB |
233 | if (!list_empty(&ctx->aff_list)) |
234 | atomic_inc(&ctx->gang->aff_sched_count); | |
aa6d5b20 | 235 | |
e9f8a0b6 CH |
236 | ctx->stats.slb_flt_base = spu->stats.slb_flt; |
237 | ctx->stats.class2_intr_base = spu->stats.class2_intr; | |
238 | ||
8b3d6663 AB |
239 | spu->ctx = ctx; |
240 | spu->flags = 0; | |
241 | ctx->spu = spu; | |
242 | ctx->ops = &spu_hw_ops; | |
243 | spu->pid = current->pid; | |
1474855d | 244 | spu->tgid = current->tgid; |
94b2a439 | 245 | spu_associate_mm(spu, ctx->owner); |
8b3d6663 AB |
246 | spu->ibox_callback = spufs_ibox_callback; |
247 | spu->wbox_callback = spufs_wbox_callback; | |
5110459f | 248 | spu->stop_callback = spufs_stop_callback; |
a33a7d73 | 249 | spu->mfc_callback = spufs_mfc_callback; |
9add11da | 250 | spu->dma_callback = spufs_dma_callback; |
8b3d6663 | 251 | mb(); |
5110459f | 252 | spu_unmap_mappings(ctx); |
8b3d6663 | 253 | spu_restore(&ctx->csa, spu); |
2a911f0b | 254 | spu->timestamp = jiffies; |
a68cf983 | 255 | spu_cpu_affinity_set(spu, raw_smp_processor_id()); |
86767277 | 256 | spu_switch_notify(spu, ctx); |
81998baf | 257 | ctx->state = SPU_STATE_RUNNABLE; |
27ec41d3 AD |
258 | |
259 | spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED); | |
8b3d6663 AB |
260 | } |
261 | ||
c5fc8d2a | 262 | /* |
486acd48 | 263 | * Must be used with the list_mutex held. |
c5fc8d2a AB |
264 | */ |
265 | static inline int sched_spu(struct spu *spu) | |
266 | { | |
486acd48 CH |
267 | BUG_ON(!mutex_is_locked(&cbe_spu_info[spu->node].list_mutex)); |
268 | ||
c5fc8d2a AB |
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; | |
486acd48 | 321 | mutex_lock(&cbe_spu_info[node].list_mutex); |
c5fc8d2a AB |
322 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { |
323 | if ((!mem_aff || spu->has_mem_affinity) && | |
486acd48 CH |
324 | sched_spu(spu)) { |
325 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
c5fc8d2a | 326 | return spu; |
486acd48 | 327 | } |
c5fc8d2a | 328 | } |
486acd48 | 329 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
c5fc8d2a AB |
330 | } |
331 | return NULL; | |
332 | } | |
333 | ||
334 | static void aff_set_ref_point_location(struct spu_gang *gang) | |
335 | { | |
336 | int mem_aff, gs, lowest_offset; | |
337 | struct spu_context *ctx; | |
338 | struct spu *tmp; | |
339 | ||
340 | mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM; | |
341 | lowest_offset = 0; | |
342 | gs = 0; | |
343 | ||
344 | list_for_each_entry(tmp, &gang->aff_list_head, aff_list) | |
345 | gs++; | |
346 | ||
347 | list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list, | |
348 | aff_list) { | |
349 | if (&ctx->aff_list == &gang->aff_list_head) | |
350 | break; | |
351 | lowest_offset = ctx->aff_offset; | |
352 | } | |
353 | ||
683e3ab2 AD |
354 | gang->aff_ref_spu = aff_ref_location(gang->aff_ref_ctx, mem_aff, gs, |
355 | lowest_offset); | |
c5fc8d2a AB |
356 | } |
357 | ||
486acd48 | 358 | static struct spu *ctx_location(struct spu *ref, int offset, int node) |
c5fc8d2a AB |
359 | { |
360 | struct spu *spu; | |
361 | ||
362 | spu = NULL; | |
363 | if (offset >= 0) { | |
364 | list_for_each_entry(spu, ref->aff_list.prev, aff_list) { | |
486acd48 | 365 | BUG_ON(spu->node != node); |
c5fc8d2a AB |
366 | if (offset == 0) |
367 | break; | |
368 | if (sched_spu(spu)) | |
369 | offset--; | |
370 | } | |
371 | } else { | |
372 | list_for_each_entry_reverse(spu, ref->aff_list.next, aff_list) { | |
486acd48 | 373 | BUG_ON(spu->node != node); |
c5fc8d2a AB |
374 | if (offset == 0) |
375 | break; | |
376 | if (sched_spu(spu)) | |
377 | offset++; | |
378 | } | |
379 | } | |
486acd48 | 380 | |
c5fc8d2a AB |
381 | return spu; |
382 | } | |
383 | ||
384 | /* | |
385 | * affinity_check is called each time a context is going to be scheduled. | |
386 | * It returns the spu ptr on which the context must run. | |
387 | */ | |
486acd48 | 388 | static int has_affinity(struct spu_context *ctx) |
c5fc8d2a | 389 | { |
486acd48 | 390 | struct spu_gang *gang = ctx->gang; |
c5fc8d2a AB |
391 | |
392 | if (list_empty(&ctx->aff_list)) | |
486acd48 CH |
393 | return 0; |
394 | ||
c5fc8d2a AB |
395 | mutex_lock(&gang->aff_mutex); |
396 | if (!gang->aff_ref_spu) { | |
397 | if (!(gang->aff_flags & AFF_MERGED)) | |
398 | aff_merge_remaining_ctxs(gang); | |
399 | if (!(gang->aff_flags & AFF_OFFSETS_SET)) | |
400 | aff_set_offsets(gang); | |
401 | aff_set_ref_point_location(gang); | |
402 | } | |
403 | mutex_unlock(&gang->aff_mutex); | |
486acd48 CH |
404 | |
405 | return gang->aff_ref_spu != NULL; | |
c5fc8d2a AB |
406 | } |
407 | ||
202557d2 CH |
408 | /** |
409 | * spu_unbind_context - unbind spu context from physical spu | |
410 | * @spu: physical spu to unbind from | |
411 | * @ctx: context to unbind | |
202557d2 | 412 | */ |
678b2ff1 | 413 | static void spu_unbind_context(struct spu *spu, struct spu_context *ctx) |
8b3d6663 | 414 | { |
a68cf983 MN |
415 | pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__, |
416 | spu->pid, spu->number, spu->node); | |
27ec41d3 | 417 | spuctx_switch_state(ctx, SPU_UTIL_SYSTEM); |
fe2f896d | 418 | |
aa6d5b20 AB |
419 | if (spu->ctx->flags & SPU_CREATE_NOSCHED) |
420 | atomic_dec(&cbe_spu_info[spu->node].reserved_spus); | |
c5fc8d2a AB |
421 | if (!list_empty(&ctx->aff_list)) |
422 | if (atomic_dec_and_test(&ctx->gang->aff_sched_count)) | |
423 | ctx->gang->aff_ref_spu = NULL; | |
86767277 | 424 | spu_switch_notify(spu, NULL); |
5110459f | 425 | spu_unmap_mappings(ctx); |
8b3d6663 | 426 | spu_save(&ctx->csa, spu); |
2a911f0b | 427 | spu->timestamp = jiffies; |
8b3d6663 AB |
428 | ctx->state = SPU_STATE_SAVED; |
429 | spu->ibox_callback = NULL; | |
430 | spu->wbox_callback = NULL; | |
5110459f | 431 | spu->stop_callback = NULL; |
a33a7d73 | 432 | spu->mfc_callback = NULL; |
9add11da | 433 | spu->dma_callback = NULL; |
94b2a439 | 434 | spu_associate_mm(spu, NULL); |
8b3d6663 | 435 | spu->pid = 0; |
1474855d | 436 | spu->tgid = 0; |
8b3d6663 | 437 | ctx->ops = &spu_backing_ops; |
2a911f0b | 438 | spu->flags = 0; |
8b3d6663 | 439 | spu->ctx = NULL; |
e9f8a0b6 CH |
440 | |
441 | ctx->stats.slb_flt += | |
442 | (spu->stats.slb_flt - ctx->stats.slb_flt_base); | |
443 | ctx->stats.class2_intr += | |
444 | (spu->stats.class2_intr - ctx->stats.class2_intr_base); | |
27ec41d3 AD |
445 | |
446 | /* This maps the underlying spu state to idle */ | |
447 | spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED); | |
448 | ctx->spu = NULL; | |
8b3d6663 AB |
449 | } |
450 | ||
079cdb61 CH |
451 | /** |
452 | * spu_add_to_rq - add a context to the runqueue | |
453 | * @ctx: context to add | |
454 | */ | |
4e0f4ed0 | 455 | static void __spu_add_to_rq(struct spu_context *ctx) |
8b3d6663 | 456 | { |
27449971 CH |
457 | /* |
458 | * Unfortunately this code path can be called from multiple threads | |
459 | * on behalf of a single context due to the way the problem state | |
460 | * mmap support works. | |
461 | * | |
462 | * Fortunately we need to wake up all these threads at the same time | |
463 | * and can simply skip the runqueue addition for every but the first | |
464 | * thread getting into this codepath. | |
465 | * | |
466 | * It's still quite hacky, and long-term we should proxy all other | |
467 | * threads through the owner thread so that spu_run is in control | |
468 | * of all the scheduling activity for a given context. | |
469 | */ | |
470 | if (list_empty(&ctx->rq)) { | |
471 | list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]); | |
472 | set_bit(ctx->prio, spu_prio->bitmap); | |
473 | if (!spu_prio->nr_waiting++) | |
474 | __mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK); | |
475 | } | |
2a911f0b | 476 | } |
5110459f | 477 | |
4e0f4ed0 | 478 | static void __spu_del_from_rq(struct spu_context *ctx) |
a475c2f4 | 479 | { |
4e0f4ed0 LB |
480 | int prio = ctx->prio; |
481 | ||
65de66f0 | 482 | if (!list_empty(&ctx->rq)) { |
c77239b8 CH |
483 | if (!--spu_prio->nr_waiting) |
484 | del_timer(&spusched_timer); | |
a475c2f4 | 485 | list_del_init(&ctx->rq); |
c77239b8 CH |
486 | |
487 | if (list_empty(&spu_prio->runq[prio])) | |
488 | clear_bit(prio, spu_prio->bitmap); | |
65de66f0 | 489 | } |
079cdb61 | 490 | } |
a68cf983 | 491 | |
079cdb61 | 492 | static void spu_prio_wait(struct spu_context *ctx) |
8b3d6663 | 493 | { |
a68cf983 | 494 | DEFINE_WAIT(wait); |
8b3d6663 | 495 | |
4e0f4ed0 | 496 | spin_lock(&spu_prio->runq_lock); |
079cdb61 | 497 | prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE); |
a68cf983 | 498 | if (!signal_pending(current)) { |
4e0f4ed0 LB |
499 | __spu_add_to_rq(ctx); |
500 | spin_unlock(&spu_prio->runq_lock); | |
650f8b02 | 501 | mutex_unlock(&ctx->state_mutex); |
a68cf983 | 502 | schedule(); |
650f8b02 | 503 | mutex_lock(&ctx->state_mutex); |
4e0f4ed0 LB |
504 | spin_lock(&spu_prio->runq_lock); |
505 | __spu_del_from_rq(ctx); | |
8b3d6663 | 506 | } |
4e0f4ed0 | 507 | spin_unlock(&spu_prio->runq_lock); |
079cdb61 CH |
508 | __set_current_state(TASK_RUNNING); |
509 | remove_wait_queue(&ctx->stop_wq, &wait); | |
8b3d6663 AB |
510 | } |
511 | ||
079cdb61 | 512 | static struct spu *spu_get_idle(struct spu_context *ctx) |
a68cf983 | 513 | { |
486acd48 CH |
514 | struct spu *spu; |
515 | int node, n; | |
516 | ||
517 | if (has_affinity(ctx)) { | |
518 | node = ctx->gang->aff_ref_spu->node; | |
a68cf983 | 519 | |
486acd48 CH |
520 | mutex_lock(&cbe_spu_info[node].list_mutex); |
521 | spu = ctx_location(ctx->gang->aff_ref_spu, ctx->aff_offset, node); | |
522 | if (spu && spu->alloc_state == SPU_FREE) | |
523 | goto found; | |
524 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
525 | return NULL; | |
526 | } | |
cbc23d3e | 527 | |
486acd48 | 528 | node = cpu_to_node(raw_smp_processor_id()); |
a68cf983 MN |
529 | for (n = 0; n < MAX_NUMNODES; n++, node++) { |
530 | node = (node < MAX_NUMNODES) ? node : 0; | |
ea1ae594 | 531 | if (!node_allowed(ctx, node)) |
a68cf983 | 532 | continue; |
486acd48 CH |
533 | |
534 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
535 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { | |
536 | if (spu->alloc_state == SPU_FREE) | |
537 | goto found; | |
538 | } | |
539 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
a68cf983 | 540 | } |
486acd48 CH |
541 | |
542 | return NULL; | |
543 | ||
544 | found: | |
545 | spu->alloc_state = SPU_USED; | |
546 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
547 | pr_debug("Got SPU %d %d\n", spu->number, spu->node); | |
548 | spu_init_channels(spu); | |
a68cf983 MN |
549 | return spu; |
550 | } | |
8b3d6663 | 551 | |
52f04fcf CH |
552 | /** |
553 | * find_victim - find a lower priority context to preempt | |
554 | * @ctx: canidate context for running | |
555 | * | |
556 | * Returns the freed physical spu to run the new context on. | |
557 | */ | |
558 | static struct spu *find_victim(struct spu_context *ctx) | |
559 | { | |
560 | struct spu_context *victim = NULL; | |
561 | struct spu *spu; | |
562 | int node, n; | |
563 | ||
564 | /* | |
565 | * Look for a possible preemption candidate on the local node first. | |
566 | * If there is no candidate look at the other nodes. This isn't | |
567 | * exactly fair, but so far the whole spu schedule tries to keep | |
568 | * a strong node affinity. We might want to fine-tune this in | |
569 | * the future. | |
570 | */ | |
571 | restart: | |
572 | node = cpu_to_node(raw_smp_processor_id()); | |
573 | for (n = 0; n < MAX_NUMNODES; n++, node++) { | |
574 | node = (node < MAX_NUMNODES) ? node : 0; | |
ea1ae594 | 575 | if (!node_allowed(ctx, node)) |
52f04fcf CH |
576 | continue; |
577 | ||
486acd48 CH |
578 | mutex_lock(&cbe_spu_info[node].list_mutex); |
579 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) { | |
52f04fcf CH |
580 | struct spu_context *tmp = spu->ctx; |
581 | ||
fe443ef2 CH |
582 | if (tmp->prio > ctx->prio && |
583 | (!victim || tmp->prio > victim->prio)) | |
52f04fcf CH |
584 | victim = spu->ctx; |
585 | } | |
486acd48 | 586 | mutex_unlock(&cbe_spu_info[node].list_mutex); |
52f04fcf CH |
587 | |
588 | if (victim) { | |
589 | /* | |
590 | * This nests ctx->state_mutex, but we always lock | |
591 | * higher priority contexts before lower priority | |
592 | * ones, so this is safe until we introduce | |
593 | * priority inheritance schemes. | |
594 | */ | |
595 | if (!mutex_trylock(&victim->state_mutex)) { | |
596 | victim = NULL; | |
597 | goto restart; | |
598 | } | |
599 | ||
600 | spu = victim->spu; | |
601 | if (!spu) { | |
602 | /* | |
603 | * This race can happen because we've dropped | |
604 | * the active list mutex. No a problem, just | |
605 | * restart the search. | |
606 | */ | |
607 | mutex_unlock(&victim->state_mutex); | |
608 | victim = NULL; | |
609 | goto restart; | |
610 | } | |
486acd48 CH |
611 | |
612 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
613 | cbe_spu_info[node].nr_active--; | |
614 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
615 | ||
52f04fcf | 616 | spu_unbind_context(spu, victim); |
e9f8a0b6 | 617 | victim->stats.invol_ctx_switch++; |
fe2f896d | 618 | spu->stats.invol_ctx_switch++; |
52f04fcf | 619 | mutex_unlock(&victim->state_mutex); |
e097b513 CH |
620 | /* |
621 | * We need to break out of the wait loop in spu_run | |
622 | * manually to ensure this context gets put on the | |
623 | * runqueue again ASAP. | |
624 | */ | |
625 | wake_up(&victim->stop_wq); | |
52f04fcf CH |
626 | return spu; |
627 | } | |
628 | } | |
629 | ||
630 | return NULL; | |
631 | } | |
632 | ||
079cdb61 CH |
633 | /** |
634 | * spu_activate - find a free spu for a context and execute it | |
635 | * @ctx: spu context to schedule | |
636 | * @flags: flags (currently ignored) | |
637 | * | |
08873095 | 638 | * Tries to find a free spu to run @ctx. If no free spu is available |
079cdb61 CH |
639 | * add the context to the runqueue so it gets woken up once an spu |
640 | * is available. | |
641 | */ | |
26bec673 | 642 | int spu_activate(struct spu_context *ctx, unsigned long flags) |
8b3d6663 | 643 | { |
079cdb61 CH |
644 | do { |
645 | struct spu *spu; | |
646 | ||
27449971 CH |
647 | /* |
648 | * If there are multiple threads waiting for a single context | |
649 | * only one actually binds the context while the others will | |
650 | * only be able to acquire the state_mutex once the context | |
651 | * already is in runnable state. | |
652 | */ | |
653 | if (ctx->spu) | |
654 | return 0; | |
655 | ||
079cdb61 | 656 | spu = spu_get_idle(ctx); |
52f04fcf CH |
657 | /* |
658 | * If this is a realtime thread we try to get it running by | |
659 | * preempting a lower priority thread. | |
660 | */ | |
fe443ef2 | 661 | if (!spu && rt_prio(ctx->prio)) |
52f04fcf | 662 | spu = find_victim(ctx); |
079cdb61 | 663 | if (spu) { |
486acd48 CH |
664 | int node = spu->node; |
665 | ||
666 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
202557d2 | 667 | spu_bind_context(spu, ctx); |
486acd48 CH |
668 | cbe_spu_info[node].nr_active++; |
669 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
079cdb61 | 670 | return 0; |
a68cf983 | 671 | } |
079cdb61 | 672 | |
50b520d4 | 673 | spu_prio_wait(ctx); |
079cdb61 CH |
674 | } while (!signal_pending(current)); |
675 | ||
676 | return -ERESTARTSYS; | |
8b3d6663 AB |
677 | } |
678 | ||
bb5db29a CH |
679 | /** |
680 | * grab_runnable_context - try to find a runnable context | |
681 | * | |
682 | * Remove the highest priority context on the runqueue and return it | |
683 | * to the caller. Returns %NULL if no runnable context was found. | |
684 | */ | |
ea1ae594 | 685 | static struct spu_context *grab_runnable_context(int prio, int node) |
bb5db29a | 686 | { |
ea1ae594 | 687 | struct spu_context *ctx; |
bb5db29a CH |
688 | int best; |
689 | ||
690 | spin_lock(&spu_prio->runq_lock); | |
7e90b749 | 691 | best = find_first_bit(spu_prio->bitmap, prio); |
ea1ae594 | 692 | while (best < prio) { |
bb5db29a CH |
693 | struct list_head *rq = &spu_prio->runq[best]; |
694 | ||
ea1ae594 CH |
695 | list_for_each_entry(ctx, rq, rq) { |
696 | /* XXX(hch): check for affinity here aswell */ | |
697 | if (__node_allowed(ctx, node)) { | |
698 | __spu_del_from_rq(ctx); | |
699 | goto found; | |
700 | } | |
701 | } | |
702 | best++; | |
bb5db29a | 703 | } |
ea1ae594 CH |
704 | ctx = NULL; |
705 | found: | |
bb5db29a | 706 | spin_unlock(&spu_prio->runq_lock); |
bb5db29a CH |
707 | return ctx; |
708 | } | |
709 | ||
710 | static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio) | |
711 | { | |
712 | struct spu *spu = ctx->spu; | |
713 | struct spu_context *new = NULL; | |
714 | ||
715 | if (spu) { | |
ea1ae594 | 716 | new = grab_runnable_context(max_prio, spu->node); |
bb5db29a | 717 | if (new || force) { |
486acd48 CH |
718 | int node = spu->node; |
719 | ||
720 | mutex_lock(&cbe_spu_info[node].list_mutex); | |
bb5db29a | 721 | spu_unbind_context(spu, ctx); |
486acd48 CH |
722 | spu->alloc_state = SPU_FREE; |
723 | cbe_spu_info[node].nr_active--; | |
724 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
725 | ||
e9f8a0b6 | 726 | ctx->stats.vol_ctx_switch++; |
fe2f896d | 727 | spu->stats.vol_ctx_switch++; |
486acd48 | 728 | |
bb5db29a CH |
729 | if (new) |
730 | wake_up(&new->stop_wq); | |
731 | } | |
732 | ||
733 | } | |
734 | ||
735 | return new != NULL; | |
736 | } | |
737 | ||
678b2ff1 CH |
738 | /** |
739 | * spu_deactivate - unbind a context from it's physical spu | |
740 | * @ctx: spu context to unbind | |
741 | * | |
742 | * Unbind @ctx from the physical spu it is running on and schedule | |
743 | * the highest priority context to run on the freed physical spu. | |
744 | */ | |
8b3d6663 AB |
745 | void spu_deactivate(struct spu_context *ctx) |
746 | { | |
bb5db29a | 747 | __spu_deactivate(ctx, 1, MAX_PRIO); |
8b3d6663 AB |
748 | } |
749 | ||
ae7b4c52 | 750 | /** |
1474855d | 751 | * spu_yield - yield a physical spu if others are waiting |
ae7b4c52 CH |
752 | * @ctx: spu context to yield |
753 | * | |
754 | * Check if there is a higher priority context waiting and if yes | |
755 | * unbind @ctx from the physical spu and schedule the highest | |
756 | * priority context to run on the freed physical spu instead. | |
757 | */ | |
8b3d6663 AB |
758 | void spu_yield(struct spu_context *ctx) |
759 | { | |
e5c0b9ec CH |
760 | if (!(ctx->flags & SPU_CREATE_NOSCHED)) { |
761 | mutex_lock(&ctx->state_mutex); | |
27ec41d3 | 762 | __spu_deactivate(ctx, 0, MAX_PRIO); |
e5c0b9ec CH |
763 | mutex_unlock(&ctx->state_mutex); |
764 | } | |
bb5db29a | 765 | } |
8b3d6663 | 766 | |
486acd48 | 767 | static noinline void spusched_tick(struct spu_context *ctx) |
bb5db29a | 768 | { |
df09cf3e CH |
769 | if (ctx->flags & SPU_CREATE_NOSCHED) |
770 | return; | |
771 | if (ctx->policy == SCHED_FIFO) | |
772 | return; | |
773 | ||
774 | if (--ctx->time_slice) | |
37901802 | 775 | return; |
bb5db29a CH |
776 | |
777 | /* | |
486acd48 | 778 | * Unfortunately list_mutex ranks outside of state_mutex, so |
37901802 CH |
779 | * we have to trylock here. If we fail give the context another |
780 | * tick and try again. | |
bb5db29a | 781 | */ |
37901802 | 782 | if (mutex_trylock(&ctx->state_mutex)) { |
7022543e | 783 | struct spu *spu = ctx->spu; |
ea1ae594 CH |
784 | struct spu_context *new; |
785 | ||
786 | new = grab_runnable_context(ctx->prio + 1, spu->node); | |
37901802 | 787 | if (new) { |
37901802 | 788 | spu_unbind_context(spu, ctx); |
e9f8a0b6 | 789 | ctx->stats.invol_ctx_switch++; |
fe2f896d | 790 | spu->stats.invol_ctx_switch++; |
486acd48 CH |
791 | spu->alloc_state = SPU_FREE; |
792 | cbe_spu_info[spu->node].nr_active--; | |
37901802 CH |
793 | wake_up(&new->stop_wq); |
794 | /* | |
795 | * We need to break out of the wait loop in | |
796 | * spu_run manually to ensure this context | |
797 | * gets put on the runqueue again ASAP. | |
798 | */ | |
799 | wake_up(&ctx->stop_wq); | |
800 | } | |
fe443ef2 | 801 | spu_set_timeslice(ctx); |
37901802 | 802 | mutex_unlock(&ctx->state_mutex); |
bb5db29a | 803 | } else { |
37901802 | 804 | ctx->time_slice++; |
8b3d6663 | 805 | } |
8b3d6663 AB |
806 | } |
807 | ||
65de66f0 CH |
808 | /** |
809 | * count_active_contexts - count nr of active tasks | |
810 | * | |
811 | * Return the number of tasks currently running or waiting to run. | |
812 | * | |
486acd48 | 813 | * Note that we don't take runq_lock / list_mutex here. Reading |
65de66f0 CH |
814 | * a single 32bit value is atomic on powerpc, and we don't care |
815 | * about memory ordering issues here. | |
816 | */ | |
817 | static unsigned long count_active_contexts(void) | |
818 | { | |
819 | int nr_active = 0, node; | |
820 | ||
821 | for (node = 0; node < MAX_NUMNODES; node++) | |
486acd48 | 822 | nr_active += cbe_spu_info[node].nr_active; |
65de66f0 CH |
823 | nr_active += spu_prio->nr_waiting; |
824 | ||
825 | return nr_active; | |
826 | } | |
827 | ||
828 | /** | |
829 | * spu_calc_load - given tick count, update the avenrun load estimates. | |
830 | * @tick: tick count | |
831 | * | |
832 | * No locking against reading these values from userspace, as for | |
833 | * the CPU loadavg code. | |
834 | */ | |
835 | static void spu_calc_load(unsigned long ticks) | |
836 | { | |
837 | unsigned long active_tasks; /* fixed-point */ | |
838 | static int count = LOAD_FREQ; | |
839 | ||
840 | count -= ticks; | |
841 | ||
842 | if (unlikely(count < 0)) { | |
843 | active_tasks = count_active_contexts() * FIXED_1; | |
844 | do { | |
845 | CALC_LOAD(spu_avenrun[0], EXP_1, active_tasks); | |
846 | CALC_LOAD(spu_avenrun[1], EXP_5, active_tasks); | |
847 | CALC_LOAD(spu_avenrun[2], EXP_15, active_tasks); | |
848 | count += LOAD_FREQ; | |
849 | } while (count < 0); | |
850 | } | |
851 | } | |
852 | ||
37901802 CH |
853 | static void spusched_wake(unsigned long data) |
854 | { | |
855 | mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK); | |
856 | wake_up_process(spusched_task); | |
65de66f0 | 857 | spu_calc_load(SPUSCHED_TICK); |
37901802 CH |
858 | } |
859 | ||
860 | static int spusched_thread(void *unused) | |
861 | { | |
486acd48 | 862 | struct spu *spu; |
37901802 CH |
863 | int node; |
864 | ||
37901802 CH |
865 | while (!kthread_should_stop()) { |
866 | set_current_state(TASK_INTERRUPTIBLE); | |
867 | schedule(); | |
868 | for (node = 0; node < MAX_NUMNODES; node++) { | |
486acd48 CH |
869 | mutex_lock(&cbe_spu_info[node].list_mutex); |
870 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) | |
871 | if (spu->ctx) | |
872 | spusched_tick(spu->ctx); | |
873 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
37901802 CH |
874 | } |
875 | } | |
876 | ||
37901802 CH |
877 | return 0; |
878 | } | |
879 | ||
65de66f0 CH |
880 | #define LOAD_INT(x) ((x) >> FSHIFT) |
881 | #define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100) | |
882 | ||
883 | static int show_spu_loadavg(struct seq_file *s, void *private) | |
884 | { | |
885 | int a, b, c; | |
886 | ||
887 | a = spu_avenrun[0] + (FIXED_1/200); | |
888 | b = spu_avenrun[1] + (FIXED_1/200); | |
889 | c = spu_avenrun[2] + (FIXED_1/200); | |
890 | ||
891 | /* | |
892 | * Note that last_pid doesn't really make much sense for the | |
893 | * SPU loadavg (it even seems very odd on the CPU side..), | |
894 | * but we include it here to have a 100% compatible interface. | |
895 | */ | |
896 | seq_printf(s, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n", | |
897 | LOAD_INT(a), LOAD_FRAC(a), | |
898 | LOAD_INT(b), LOAD_FRAC(b), | |
899 | LOAD_INT(c), LOAD_FRAC(c), | |
900 | count_active_contexts(), | |
901 | atomic_read(&nr_spu_contexts), | |
902 | current->nsproxy->pid_ns->last_pid); | |
903 | return 0; | |
904 | } | |
905 | ||
906 | static int spu_loadavg_open(struct inode *inode, struct file *file) | |
907 | { | |
908 | return single_open(file, show_spu_loadavg, NULL); | |
909 | } | |
910 | ||
911 | static const struct file_operations spu_loadavg_fops = { | |
912 | .open = spu_loadavg_open, | |
913 | .read = seq_read, | |
914 | .llseek = seq_lseek, | |
915 | .release = single_release, | |
916 | }; | |
917 | ||
8b3d6663 AB |
918 | int __init spu_sched_init(void) |
919 | { | |
65de66f0 CH |
920 | struct proc_dir_entry *entry; |
921 | int err = -ENOMEM, i; | |
8b3d6663 | 922 | |
a68cf983 | 923 | spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL); |
37901802 | 924 | if (!spu_prio) |
65de66f0 | 925 | goto out; |
37901802 | 926 | |
8b3d6663 | 927 | for (i = 0; i < MAX_PRIO; i++) { |
079cdb61 | 928 | INIT_LIST_HEAD(&spu_prio->runq[i]); |
a68cf983 | 929 | __clear_bit(i, spu_prio->bitmap); |
8b3d6663 | 930 | } |
079cdb61 | 931 | spin_lock_init(&spu_prio->runq_lock); |
37901802 | 932 | |
c77239b8 CH |
933 | setup_timer(&spusched_timer, spusched_wake, 0); |
934 | ||
37901802 CH |
935 | spusched_task = kthread_run(spusched_thread, NULL, "spusched"); |
936 | if (IS_ERR(spusched_task)) { | |
65de66f0 CH |
937 | err = PTR_ERR(spusched_task); |
938 | goto out_free_spu_prio; | |
37901802 | 939 | } |
f3f59bec | 940 | |
65de66f0 CH |
941 | entry = create_proc_entry("spu_loadavg", 0, NULL); |
942 | if (!entry) | |
943 | goto out_stop_kthread; | |
944 | entry->proc_fops = &spu_loadavg_fops; | |
945 | ||
f3f59bec JK |
946 | pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n", |
947 | SPUSCHED_TICK, MIN_SPU_TIMESLICE, DEF_SPU_TIMESLICE); | |
8b3d6663 | 948 | return 0; |
37901802 | 949 | |
65de66f0 CH |
950 | out_stop_kthread: |
951 | kthread_stop(spusched_task); | |
952 | out_free_spu_prio: | |
953 | kfree(spu_prio); | |
954 | out: | |
955 | return err; | |
8b3d6663 AB |
956 | } |
957 | ||
d1450317 | 958 | void spu_sched_exit(void) |
8b3d6663 | 959 | { |
486acd48 | 960 | struct spu *spu; |
a68cf983 MN |
961 | int node; |
962 | ||
65de66f0 CH |
963 | remove_proc_entry("spu_loadavg", NULL); |
964 | ||
c77239b8 | 965 | del_timer_sync(&spusched_timer); |
37901802 CH |
966 | kthread_stop(spusched_task); |
967 | ||
a68cf983 | 968 | for (node = 0; node < MAX_NUMNODES; node++) { |
486acd48 CH |
969 | mutex_lock(&cbe_spu_info[node].list_mutex); |
970 | list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) | |
971 | if (spu->alloc_state != SPU_FREE) | |
972 | spu->alloc_state = SPU_FREE; | |
973 | mutex_unlock(&cbe_spu_info[node].list_mutex); | |
8b3d6663 | 974 | } |
a68cf983 | 975 | kfree(spu_prio); |
8b3d6663 | 976 | } |