arch/powerpc/platforms/cell/spufs/sched.c

   1 /* sched.c - SPU scheduler.
   2  *
   3  * Copyright (C) IBM 2005
   4  * Author: Mark Nutter <mnutter@us.ibm.com>
   5  *
   6  * 2006-03-31   NUMA domains added.
   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
  23 #undef DEBUG
  24
  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>
  33 #include <linux/smp_lock.h>
  34 #include <linux/stddef.h>
  35 #include <linux/unistd.h>
  36 #include <linux/numa.h>
  37 #include <linux/mutex.h>
  38 #include <linux/notifier.h>
  39
  40 #include <asm/io.h>
  41 #include <asm/mmu_context.h>
  42 #include <asm/spu.h>
  43 #include <asm/spu_csa.h>
  44 #include <asm/spu_priv1.h>
  45 #include "spufs.h"
  46
  47 #define SPU_TIMESLICE   (HZ)
  48
  49 struct spu_prio_array {
  50         DECLARE_BITMAP(bitmap, MAX_PRIO);
  51         struct list_head runq[MAX_PRIO];
  52         spinlock_t runq_lock;
  53         struct list_head active_list[MAX_NUMNODES];
  54         struct mutex active_mutex[MAX_NUMNODES];
  55 };
  56
  57 static struct spu_prio_array *spu_prio;
  58 static struct workqueue_struct *spu_sched_wq;
  59
  60 static inline int node_allowed(int node)
  61 {
  62         cpumask_t mask;
  63
  64         if (!nr_cpus_node(node))
  65                 return 0;
  66         mask = node_to_cpumask(node);
  67         if (!cpus_intersects(mask, current->cpus_allowed))
  68                 return 0;
  69         return 1;
  70 }
  71
  72 void spu_start_tick(struct spu_context *ctx)
  73 {
  74         if (ctx->policy == SCHED_RR) {
  75                 /*
  76                  * Make sure the exiting bit is cleared.
  77                  */
  78                 clear_bit(SPU_SCHED_EXITING, &ctx->sched_flags);
  79                 mb();
  80                 queue_delayed_work(spu_sched_wq, &ctx->sched_work, SPU_TIMESLICE);
  81         }
  82 }
  83
  84 void spu_stop_tick(struct spu_context *ctx)
  85 {
  86         if (ctx->policy == SCHED_RR) {
  87                 /*
  88                  * While the work can be rearming normally setting this flag
  89                  * makes sure it does not rearm itself anymore.
  90                  */
  91                 set_bit(SPU_SCHED_EXITING, &ctx->sched_flags);
  92                 mb();
  93                 cancel_delayed_work(&ctx->sched_work);
  94         }
  95 }
  96
  97 void spu_sched_tick(struct work_struct *work)
  98 {
  99         struct spu_context *ctx =
 100                 container_of(work, struct spu_context, sched_work.work);
 101         struct spu *spu;
 102         int preempted = 0;
 103
 104         /*
 105          * If this context is being stopped avoid rescheduling from the
 106          * scheduler tick because we would block on the state_mutex.
 107          * The caller will yield the spu later on anyway.
 108          */
 109         if (test_bit(SPU_SCHED_EXITING, &ctx->sched_flags))
 110                 return;
 111
 112         mutex_lock(&ctx->state_mutex);
 113         spu = ctx->spu;
 114         if (spu) {
 115                 int best = sched_find_first_bit(spu_prio->bitmap);
 116                 if (best <= ctx->prio) {
 117                         spu_deactivate(ctx);
 118                         preempted = 1;
 119                 }
 120         }
 121         mutex_unlock(&ctx->state_mutex);
 122
 123         if (preempted) {
 124                 /*
 125                  * We need to break out of the wait loop in spu_run manually
 126                  * to ensure this context gets put on the runqueue again
 127                  * ASAP.
 128                  */
 129                 wake_up(&ctx->stop_wq);
 130         } else
 131                 spu_start_tick(ctx);
 132 }
 133
 134 /**
 135  * spu_add_to_active_list - add spu to active list
 136  * @spu:        spu to add to the active list
 137  */
 138 static void spu_add_to_active_list(struct spu *spu)
 139 {
 140         mutex_lock(&spu_prio->active_mutex[spu->node]);
 141         list_add_tail(&spu->list, &spu_prio->active_list[spu->node]);
 142         mutex_unlock(&spu_prio->active_mutex[spu->node]);
 143 }
 144
 145 /**
 146  * spu_remove_from_active_list - remove spu from active list
 147  * @spu:       spu to remove from the active list
 148  */
 149 static void spu_remove_from_active_list(struct spu *spu)
 150 {
 151         int node = spu->node;
 152
 153         mutex_lock(&spu_prio->active_mutex[node]);
 154         list_del_init(&spu->list);
 155         mutex_unlock(&spu_prio->active_mutex[node]);
 156 }
 157
 158 static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier);
 159
 160 static void spu_switch_notify(struct spu *spu, struct spu_context *ctx)
 161 {
 162         blocking_notifier_call_chain(&spu_switch_notifier,
 163                             ctx ? ctx->object_id : 0, spu);
 164 }
 165
 166 int spu_switch_event_register(struct notifier_block * n)
 167 {
 168         return blocking_notifier_chain_register(&spu_switch_notifier, n);
 169 }
 170
 171 int spu_switch_event_unregister(struct notifier_block * n)
 172 {
 173         return blocking_notifier_chain_unregister(&spu_switch_notifier, n);
 174 }
 175
 176 /**
 177  * spu_bind_context - bind spu context to physical spu
 178  * @spu:        physical spu to bind to
 179  * @ctx:        context to bind
 180  */
 181 static void spu_bind_context(struct spu *spu, struct spu_context *ctx)
 182 {
 183         pr_debug("%s: pid=%d SPU=%d NODE=%d\n", __FUNCTION__, current->pid,
 184                  spu->number, spu->node);
 185         spu->ctx = ctx;
 186         spu->flags = 0;
 187         ctx->spu = spu;
 188         ctx->ops = &spu_hw_ops;
 189         spu->pid = current->pid;
 190         spu_associate_mm(spu, ctx->owner);
 191         spu->ibox_callback = spufs_ibox_callback;
 192         spu->wbox_callback = spufs_wbox_callback;
 193         spu->stop_callback = spufs_stop_callback;
 194         spu->mfc_callback = spufs_mfc_callback;
 195         spu->dma_callback = spufs_dma_callback;
 196         mb();
 197         spu_unmap_mappings(ctx);
 198         spu_restore(&ctx->csa, spu);
 199         spu->timestamp = jiffies;
 200         spu_cpu_affinity_set(spu, raw_smp_processor_id());
 201         spu_switch_notify(spu, ctx);
 202         spu_add_to_active_list(spu);
 203         ctx->state = SPU_STATE_RUNNABLE;
 204 }
 205
 206 /**
 207  * spu_unbind_context - unbind spu context from physical spu
 208  * @spu:        physical spu to unbind from
 209  * @ctx:        context to unbind
 210  */
 211 static void spu_unbind_context(struct spu *spu, struct spu_context *ctx)
 212 {
 213         pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__,
 214                  spu->pid, spu->number, spu->node);
 215
 216         spu_remove_from_active_list(spu);
 217         spu_switch_notify(spu, NULL);
 218         spu_unmap_mappings(ctx);
 219         spu_save(&ctx->csa, spu);
 220         spu->timestamp = jiffies;
 221         ctx->state = SPU_STATE_SAVED;
 222         spu->ibox_callback = NULL;
 223         spu->wbox_callback = NULL;
 224         spu->stop_callback = NULL;
 225         spu->mfc_callback = NULL;
 226         spu->dma_callback = NULL;
 227         spu_associate_mm(spu, NULL);
 228         spu->pid = 0;
 229         ctx->ops = &spu_backing_ops;
 230         ctx->spu = NULL;
 231         spu->flags = 0;
 232         spu->ctx = NULL;
 233 }
 234
 235 /**
 236  * spu_add_to_rq - add a context to the runqueue
 237  * @ctx:       context to add
 238  */
 239 static void __spu_add_to_rq(struct spu_context *ctx)
 240 {
 241         int prio = ctx->prio;
 242
 243         list_add_tail(&ctx->rq, &spu_prio->runq[prio]);
 244         set_bit(prio, spu_prio->bitmap);
 245 }
 246
 247 static void __spu_del_from_rq(struct spu_context *ctx)
 248 {
 249         int prio = ctx->prio;
 250
 251         if (!list_empty(&ctx->rq))
 252                 list_del_init(&ctx->rq);
 253         if (list_empty(&spu_prio->runq[prio]))
 254                 clear_bit(prio, spu_prio->bitmap);
 255 }
 256
 257 static void spu_prio_wait(struct spu_context *ctx)
 258 {
 259         DEFINE_WAIT(wait);
 260
 261         spin_lock(&spu_prio->runq_lock);
 262         prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE);
 263         if (!signal_pending(current)) {
 264                 __spu_add_to_rq(ctx);
 265                 spin_unlock(&spu_prio->runq_lock);
 266                 mutex_unlock(&ctx->state_mutex);
 267                 schedule();
 268                 mutex_lock(&ctx->state_mutex);
 269                 spin_lock(&spu_prio->runq_lock);
 270                 __spu_del_from_rq(ctx);
 271         }
 272         spin_unlock(&spu_prio->runq_lock);
 273         __set_current_state(TASK_RUNNING);
 274         remove_wait_queue(&ctx->stop_wq, &wait);
 275 }
 276
 277 /**
 278  * spu_reschedule - try to find a runnable context for a spu
 279  * @spu:       spu available
 280  *
 281  * This function is called whenever a spu becomes idle.  It looks for the
 282  * most suitable runnable spu context and schedules it for execution.
 283  */
 284 static void spu_reschedule(struct spu *spu)
 285 {
 286         int best;
 287
 288         spu_free(spu);
 289
 290         spin_lock(&spu_prio->runq_lock);
 291         best = sched_find_first_bit(spu_prio->bitmap);
 292         if (best < MAX_PRIO) {
 293                 struct list_head *rq = &spu_prio->runq[best];
 294                 struct spu_context *ctx;
 295
 296                 BUG_ON(list_empty(rq));
 297
 298                 ctx = list_entry(rq->next, struct spu_context, rq);
 299                 __spu_del_from_rq(ctx);
 300                 wake_up(&ctx->stop_wq);
 301         }
 302         spin_unlock(&spu_prio->runq_lock);
 303 }
 304
 305 static struct spu *spu_get_idle(struct spu_context *ctx)
 306 {
 307         struct spu *spu = NULL;
 308         int node = cpu_to_node(raw_smp_processor_id());
 309         int n;
 310
 311         for (n = 0; n < MAX_NUMNODES; n++, node++) {
 312                 node = (node < MAX_NUMNODES) ? node : 0;
 313                 if (!node_allowed(node))
 314                         continue;
 315                 spu = spu_alloc_node(node);
 316                 if (spu)
 317                         break;
 318         }
 319         return spu;
 320 }
 321
 322 /**
 323  * find_victim - find a lower priority context to preempt
 324  * @ctx:        canidate context for running
 325  *
 326  * Returns the freed physical spu to run the new context on.
 327  */
 328 static struct spu *find_victim(struct spu_context *ctx)
 329 {
 330         struct spu_context *victim = NULL;
 331         struct spu *spu;
 332         int node, n;
 333
 334         /*
 335          * Look for a possible preemption candidate on the local node first.
 336          * If there is no candidate look at the other nodes.  This isn't
 337          * exactly fair, but so far the whole spu schedule tries to keep
 338          * a strong node affinity.  We might want to fine-tune this in
 339          * the future.
 340          */
 341  restart:
 342         node = cpu_to_node(raw_smp_processor_id());
 343         for (n = 0; n < MAX_NUMNODES; n++, node++) {
 344                 node = (node < MAX_NUMNODES) ? node : 0;
 345                 if (!node_allowed(node))
 346                         continue;
 347
 348                 mutex_lock(&spu_prio->active_mutex[node]);
 349                 list_for_each_entry(spu, &spu_prio->active_list[node], list) {
 350                         struct spu_context *tmp = spu->ctx;
 351
 352                         if (tmp->rt_priority < ctx->rt_priority &&
 353                             (!victim || tmp->rt_priority < victim->rt_priority))
 354                                 victim = spu->ctx;
 355                 }
 356                 mutex_unlock(&spu_prio->active_mutex[node]);
 357
 358                 if (victim) {
 359                         /*
 360                          * This nests ctx->state_mutex, but we always lock
 361                          * higher priority contexts before lower priority
 362                          * ones, so this is safe until we introduce
 363                          * priority inheritance schemes.
 364                          */
 365                         if (!mutex_trylock(&victim->state_mutex)) {
 366                                 victim = NULL;
 367                                 goto restart;
 368                         }
 369
 370                         spu = victim->spu;
 371                         if (!spu) {
 372                                 /*
 373                                  * This race can happen because we've dropped
 374                                  * the active list mutex.  No a problem, just
 375                                  * restart the search.
 376                                  */
 377                                 mutex_unlock(&victim->state_mutex);
 378                                 victim = NULL;
 379                                 goto restart;
 380                         }
 381                         spu_unbind_context(spu, victim);
 382                         mutex_unlock(&victim->state_mutex);
 383                         /*
 384                          * We need to break out of the wait loop in spu_run
 385                          * manually to ensure this context gets put on the
 386                          * runqueue again ASAP.
 387                          */
 388                         wake_up(&victim->stop_wq);
 389                         return spu;
 390                 }
 391         }
 392
 393         return NULL;
 394 }
 395
 396 /**
 397  * spu_activate - find a free spu for a context and execute it
 398  * @ctx:        spu context to schedule
 399  * @flags:      flags (currently ignored)
 400  *
 401  * Tries to find a free spu to run @ctx.  If no free spu is available
 402  * add the context to the runqueue so it gets woken up once an spu
 403  * is available.
 404  */
 405 int spu_activate(struct spu_context *ctx, unsigned long flags)
 406 {
 407
 408         if (ctx->spu)
 409                 return 0;
 410
 411         do {
 412                 struct spu *spu;
 413
 414                 spu = spu_get_idle(ctx);
 415                 /*
 416                  * If this is a realtime thread we try to get it running by
 417                  * preempting a lower priority thread.
 418                  */
 419                 if (!spu && ctx->rt_priority)
 420                         spu = find_victim(ctx);
 421                 if (spu) {
 422                         spu_bind_context(spu, ctx);
 423                         return 0;
 424                 }
 425
 426                 spu_prio_wait(ctx);
 427         } while (!signal_pending(current));
 428
 429         return -ERESTARTSYS;
 430 }
 431
 432 /**
 433  * spu_deactivate - unbind a context from it's physical spu
 434  * @ctx:        spu context to unbind
 435  *
 436  * Unbind @ctx from the physical spu it is running on and schedule
 437  * the highest priority context to run on the freed physical spu.
 438  */
 439 void spu_deactivate(struct spu_context *ctx)
 440 {
 441         struct spu *spu = ctx->spu;
 442
 443         if (spu) {
 444                 spu_unbind_context(spu, ctx);
 445                 spu_reschedule(spu);
 446         }
 447 }
 448
 449 /**
 450  * spu_yield -  yield a physical spu if others are waiting
 451  * @ctx:        spu context to yield
 452  *
 453  * Check if there is a higher priority context waiting and if yes
 454  * unbind @ctx from the physical spu and schedule the highest
 455  * priority context to run on the freed physical spu instead.
 456  */
 457 void spu_yield(struct spu_context *ctx)
 458 {
 459         struct spu *spu;
 460
 461         if (mutex_trylock(&ctx->state_mutex)) {
 462                 if ((spu = ctx->spu) != NULL) {
 463                         int best = sched_find_first_bit(spu_prio->bitmap);
 464                         if (best < MAX_PRIO) {
 465                                 pr_debug("%s: yielding SPU %d NODE %d\n",
 466                                          __FUNCTION__, spu->number, spu->node);
 467                                 spu_deactivate(ctx);
 468                         }
 469                 }
 470                 mutex_unlock(&ctx->state_mutex);
 471         }
 472 }
 473
 474 int __init spu_sched_init(void)
 475 {
 476         int i;
 477
 478         spu_sched_wq = create_singlethread_workqueue("spusched");
 479         if (!spu_sched_wq)
 480                 return 1;
 481
 482         spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
 483         if (!spu_prio) {
 484                 printk(KERN_WARNING "%s: Unable to allocate priority queue.\n",
 485                        __FUNCTION__);
 486                        destroy_workqueue(spu_sched_wq);
 487                 return 1;
 488         }
 489         for (i = 0; i < MAX_PRIO; i++) {
 490                 INIT_LIST_HEAD(&spu_prio->runq[i]);
 491                 __clear_bit(i, spu_prio->bitmap);
 492         }
 493         __set_bit(MAX_PRIO, spu_prio->bitmap);
 494         for (i = 0; i < MAX_NUMNODES; i++) {
 495                 mutex_init(&spu_prio->active_mutex[i]);
 496                 INIT_LIST_HEAD(&spu_prio->active_list[i]);
 497         }
 498         spin_lock_init(&spu_prio->runq_lock);
 499         return 0;
 500 }
 501
 502 void __exit spu_sched_exit(void)
 503 {
 504         struct spu *spu, *tmp;
 505         int node;
 506
 507         for (node = 0; node < MAX_NUMNODES; node++) {
 508                 mutex_lock(&spu_prio->active_mutex[node]);
 509                 list_for_each_entry_safe(spu, tmp, &spu_prio->active_list[node],
 510                                          list) {
 511                         list_del_init(&spu->list);
 512                         spu_free(spu);
 513                 }
 514                 mutex_unlock(&spu_prio->active_mutex[node]);
 515         }
 516         kfree(spu_prio);
 517         destroy_workqueue(spu_sched_wq);
 518 }