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_MIN_TIMESLICE       (100 * HZ / 1000)
  48
  49 #define SPU_BITMAP_SIZE (((MAX_PRIO+BITS_PER_LONG)/BITS_PER_LONG)+1)
  50 struct spu_prio_array {
  51         unsigned long bitmap[SPU_BITMAP_SIZE];
  52         wait_queue_head_t waitq[MAX_PRIO];
  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
  59 static inline int node_allowed(int node)
  60 {
  61         cpumask_t mask;
  62
  63         if (!nr_cpus_node(node))
  64                 return 0;
  65         mask = node_to_cpumask(node);
  66         if (!cpus_intersects(mask, current->cpus_allowed))
  67                 return 0;
  68         return 1;
  69 }
  70
  71 static inline void mm_needs_global_tlbie(struct mm_struct *mm)
  72 {
  73         int nr = (NR_CPUS > 1) ? NR_CPUS : NR_CPUS + 1;
  74
  75         /* Global TLBIE broadcast required with SPEs. */
  76         __cpus_setall(&mm->cpu_vm_mask, nr);
  77 }
  78
  79 static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier);
  80
  81 static void spu_switch_notify(struct spu *spu, struct spu_context *ctx)
  82 {
  83         blocking_notifier_call_chain(&spu_switch_notifier,
  84                             ctx ? ctx->object_id : 0, spu);
  85 }
  86
  87 int spu_switch_event_register(struct notifier_block * n)
  88 {
  89         return blocking_notifier_chain_register(&spu_switch_notifier, n);
  90 }
  91
  92 int spu_switch_event_unregister(struct notifier_block * n)
  93 {
  94         return blocking_notifier_chain_unregister(&spu_switch_notifier, n);
  95 }
  96
  97
  98 static inline void bind_context(struct spu *spu, struct spu_context *ctx)
  99 {
 100         pr_debug("%s: pid=%d SPU=%d NODE=%d\n", __FUNCTION__, current->pid,
 101                  spu->number, spu->node);
 102         spu->ctx = ctx;
 103         spu->flags = 0;
 104         ctx->spu = spu;
 105         ctx->ops = &spu_hw_ops;
 106         spu->pid = current->pid;
 107         spu->prio = current->prio;
 108         spu->mm = ctx->owner;
 109         mm_needs_global_tlbie(spu->mm);
 110         spu->ibox_callback = spufs_ibox_callback;
 111         spu->wbox_callback = spufs_wbox_callback;
 112         spu->stop_callback = spufs_stop_callback;
 113         spu->mfc_callback = spufs_mfc_callback;
 114         spu->dma_callback = spufs_dma_callback;
 115         mb();
 116         spu_unmap_mappings(ctx);
 117         spu_restore(&ctx->csa, spu);
 118         spu->timestamp = jiffies;
 119         spu_cpu_affinity_set(spu, raw_smp_processor_id());
 120         spu_switch_notify(spu, ctx);
 121
 122         ctx->state = SPU_STATE_RUNNABLE;
 123 }
 124
 125 static inline void unbind_context(struct spu *spu, struct spu_context *ctx)
 126 {
 127         pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__,
 128                  spu->pid, spu->number, spu->node);
 129         spu_switch_notify(spu, NULL);
 130         spu_unmap_mappings(ctx);
 131         spu_save(&ctx->csa, spu);
 132         spu->timestamp = jiffies;
 133         ctx->state = SPU_STATE_SAVED;
 134         spu->ibox_callback = NULL;
 135         spu->wbox_callback = NULL;
 136         spu->stop_callback = NULL;
 137         spu->mfc_callback = NULL;
 138         spu->dma_callback = NULL;
 139         spu->mm = NULL;
 140         spu->pid = 0;
 141         spu->prio = MAX_PRIO;
 142         ctx->ops = &spu_backing_ops;
 143         ctx->spu = NULL;
 144         spu->flags = 0;
 145         spu->ctx = NULL;
 146 }
 147
 148 static inline void spu_add_wq(wait_queue_head_t * wq, wait_queue_t * wait,
 149                               int prio)
 150 {
 151         prepare_to_wait_exclusive(wq, wait, TASK_INTERRUPTIBLE);
 152         set_bit(prio, spu_prio->bitmap);
 153 }
 154
 155 static inline void spu_del_wq(wait_queue_head_t * wq, wait_queue_t * wait,
 156                               int prio)
 157 {
 158         u64 flags;
 159
 160         __set_current_state(TASK_RUNNING);
 161
 162         spin_lock_irqsave(&wq->lock, flags);
 163
 164         remove_wait_queue_locked(wq, wait);
 165         if (list_empty(&wq->task_list))
 166                 clear_bit(prio, spu_prio->bitmap);
 167
 168         spin_unlock_irqrestore(&wq->lock, flags);
 169 }
 170
 171 static void spu_prio_wait(struct spu_context *ctx, u64 flags)
 172 {
 173         int prio = current->prio;
 174         wait_queue_head_t *wq = &spu_prio->waitq[prio];
 175         DEFINE_WAIT(wait);
 176
 177         if (ctx->spu)
 178                 return;
 179
 180         spu_add_wq(wq, &wait, prio);
 181
 182         if (!signal_pending(current)) {
 183                 up_write(&ctx->state_sema);
 184                 pr_debug("%s: pid=%d prio=%d\n", __FUNCTION__,
 185                          current->pid, current->prio);
 186                 schedule();
 187                 down_write(&ctx->state_sema);
 188         }
 189
 190         spu_del_wq(wq, &wait, prio);
 191 }
 192
 193 static void spu_prio_wakeup(void)
 194 {
 195         int best = sched_find_first_bit(spu_prio->bitmap);
 196         if (best < MAX_PRIO) {
 197                 wait_queue_head_t *wq = &spu_prio->waitq[best];
 198                 wake_up_interruptible_nr(wq, 1);
 199         }
 200 }
 201
 202 static int get_active_spu(struct spu *spu)
 203 {
 204         int node = spu->node;
 205         struct spu *tmp;
 206         int rc = 0;
 207
 208         mutex_lock(&spu_prio->active_mutex[node]);
 209         list_for_each_entry(tmp, &spu_prio->active_list[node], list) {
 210                 if (tmp == spu) {
 211                         list_del_init(&spu->list);
 212                         rc = 1;
 213                         break;
 214                 }
 215         }
 216         mutex_unlock(&spu_prio->active_mutex[node]);
 217         return rc;
 218 }
 219
 220 static void put_active_spu(struct spu *spu)
 221 {
 222         int node = spu->node;
 223
 224         mutex_lock(&spu_prio->active_mutex[node]);
 225         list_add_tail(&spu->list, &spu_prio->active_list[node]);
 226         mutex_unlock(&spu_prio->active_mutex[node]);
 227 }
 228
 229 static struct spu *spu_get_idle(struct spu_context *ctx, u64 flags)
 230 {
 231         struct spu *spu = NULL;
 232         int node = cpu_to_node(raw_smp_processor_id());
 233         int n;
 234
 235         for (n = 0; n < MAX_NUMNODES; n++, node++) {
 236                 node = (node < MAX_NUMNODES) ? node : 0;
 237                 if (!node_allowed(node))
 238                         continue;
 239                 spu = spu_alloc_node(node);
 240                 if (spu)
 241                         break;
 242         }
 243         return spu;
 244 }
 245
 246 static inline struct spu *spu_get(struct spu_context *ctx, u64 flags)
 247 {
 248         /* Future: spu_get_idle() if possible,
 249          * otherwise try to preempt an active
 250          * context.
 251          */
 252         return spu_get_idle(ctx, flags);
 253 }
 254
 255 /* The three externally callable interfaces
 256  * for the scheduler begin here.
 257  *
 258  *      spu_activate    - bind a context to SPU, waiting as needed.
 259  *      spu_deactivate  - unbind a context from its SPU.
 260  *      spu_yield       - yield an SPU if others are waiting.
 261  */
 262
 263 int spu_activate(struct spu_context *ctx, u64 flags)
 264 {
 265         struct spu *spu;
 266         int ret = 0;
 267
 268         for (;;) {
 269                 if (ctx->spu)
 270                         return 0;
 271                 spu = spu_get(ctx, flags);
 272                 if (spu != NULL) {
 273                         if (ctx->spu != NULL) {
 274                                 spu_free(spu);
 275                                 spu_prio_wakeup();
 276                                 break;
 277                         }
 278                         bind_context(spu, ctx);
 279                         put_active_spu(spu);
 280                         break;
 281                 }
 282                 spu_prio_wait(ctx, flags);
 283                 if (signal_pending(current)) {
 284                         ret = -ERESTARTSYS;
 285                         spu_prio_wakeup();
 286                         break;
 287                 }
 288         }
 289         return ret;
 290 }
 291
 292 void spu_deactivate(struct spu_context *ctx)
 293 {
 294         struct spu *spu;
 295         int needs_idle;
 296
 297         spu = ctx->spu;
 298         if (!spu)
 299                 return;
 300         needs_idle = get_active_spu(spu);
 301         unbind_context(spu, ctx);
 302         if (needs_idle) {
 303                 spu_free(spu);
 304                 spu_prio_wakeup();
 305         }
 306 }
 307
 308 void spu_yield(struct spu_context *ctx)
 309 {
 310         struct spu *spu;
 311         int need_yield = 0;
 312
 313         if (down_write_trylock(&ctx->state_sema)) {
 314                 if ((spu = ctx->spu) != NULL) {
 315                         int best = sched_find_first_bit(spu_prio->bitmap);
 316                         if (best < MAX_PRIO) {
 317                                 pr_debug("%s: yielding SPU %d NODE %d\n",
 318                                          __FUNCTION__, spu->number, spu->node);
 319                                 spu_deactivate(ctx);
 320                                 need_yield = 1;
 321                         } else {
 322                                 spu->prio = MAX_PRIO;
 323                         }
 324                 }
 325                 up_write(&ctx->state_sema);
 326         }
 327         if (unlikely(need_yield))
 328                 yield();
 329 }
 330
 331 int __init spu_sched_init(void)
 332 {
 333         int i;
 334
 335         spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
 336         if (!spu_prio) {
 337                 printk(KERN_WARNING "%s: Unable to allocate priority queue.\n",
 338                        __FUNCTION__);
 339                 return 1;
 340         }
 341         for (i = 0; i < MAX_PRIO; i++) {
 342                 init_waitqueue_head(&spu_prio->waitq[i]);
 343                 __clear_bit(i, spu_prio->bitmap);
 344         }
 345         __set_bit(MAX_PRIO, spu_prio->bitmap);
 346         for (i = 0; i < MAX_NUMNODES; i++) {
 347                 mutex_init(&spu_prio->active_mutex[i]);
 348                 INIT_LIST_HEAD(&spu_prio->active_list[i]);
 349         }
 350         return 0;
 351 }
 352
 353 void __exit spu_sched_exit(void)
 354 {
 355         struct spu *spu, *tmp;
 356         int node;
 357
 358         for (node = 0; node < MAX_NUMNODES; node++) {
 359                 mutex_lock(&spu_prio->active_mutex[node]);
 360                 list_for_each_entry_safe(spu, tmp, &spu_prio->active_list[node],
 361                                          list) {
 362                         list_del_init(&spu->list);
 363                         spu_free(spu);
 364                 }
 365                 mutex_unlock(&spu_prio->active_mutex[node]);
 366         }
 367         kfree(spu_prio);
 368 }