git.oblomov.eu Git - linux-2.6/blob - 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
  39 #include <asm/io.h>
  40 #include <asm/mmu_context.h>
  41 #include <asm/spu.h>
  42 #include <asm/spu_csa.h>
  43 #include <asm/spu_priv1.h>
  44 #include "spufs.h"
  45
  46 #define SPU_MIN_TIMESLICE       (100 * HZ / 1000)
  47
  48 #define SPU_BITMAP_SIZE (((MAX_PRIO+BITS_PER_LONG)/BITS_PER_LONG)+1)
  49 struct spu_prio_array {
  50         unsigned long bitmap[SPU_BITMAP_SIZE];
  51         wait_queue_head_t waitq[MAX_PRIO];
  52         struct list_head active_list[MAX_NUMNODES];
  53         struct mutex active_mutex[MAX_NUMNODES];
  54 };
  55
  56 static struct spu_prio_array *spu_prio;
  57
  58 static inline int node_allowed(int node)
  59 {
  60         cpumask_t mask;
  61
  62         if (!nr_cpus_node(node))
  63                 return 0;
  64         mask = node_to_cpumask(node);
  65         if (!cpus_intersects(mask, current->cpus_allowed))
  66                 return 0;
  67         return 1;
  68 }
  69
  70 static inline void mm_needs_global_tlbie(struct mm_struct *mm)
  71 {
  72         int nr = (NR_CPUS > 1) ? NR_CPUS : NR_CPUS + 1;
  73
  74         /* Global TLBIE broadcast required with SPEs. */
  75         __cpus_setall(&mm->cpu_vm_mask, nr);
  76 }
  77
  78 static inline void bind_context(struct spu *spu, struct spu_context *ctx)
  79 {
  80         pr_debug("%s: pid=%d SPU=%d NODE=%d\n", __FUNCTION__, current->pid,
  81                  spu->number, spu->node);
  82         spu->ctx = ctx;
  83         spu->flags = 0;
  84         ctx->spu = spu;
  85         ctx->ops = &spu_hw_ops;
  86         spu->pid = current->pid;
  87         spu->prio = current->prio;
  88         spu->mm = ctx->owner;
  89         mm_needs_global_tlbie(spu->mm);
  90         spu->ibox_callback = spufs_ibox_callback;
  91         spu->wbox_callback = spufs_wbox_callback;
  92         spu->stop_callback = spufs_stop_callback;
  93         spu->mfc_callback = spufs_mfc_callback;
  94         spu->dma_callback = spufs_dma_callback;
  95         mb();
  96         spu_unmap_mappings(ctx);
  97         spu_restore(&ctx->csa, spu);
  98         spu->timestamp = jiffies;
  99         spu_cpu_affinity_set(spu, raw_smp_processor_id());
 100 }
 101
 102 static inline void unbind_context(struct spu *spu, struct spu_context *ctx)
 103 {
 104         pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__,
 105                  spu->pid, spu->number, spu->node);
 106         spu_unmap_mappings(ctx);
 107         spu_save(&ctx->csa, spu);
 108         spu->timestamp = jiffies;
 109         ctx->state = SPU_STATE_SAVED;
 110         spu->ibox_callback = NULL;
 111         spu->wbox_callback = NULL;
 112         spu->stop_callback = NULL;
 113         spu->mfc_callback = NULL;
 114         spu->dma_callback = NULL;
 115         spu->mm = NULL;
 116         spu->pid = 0;
 117         spu->prio = MAX_PRIO;
 118         ctx->ops = &spu_backing_ops;
 119         ctx->spu = NULL;
 120         spu->flags = 0;
 121         spu->ctx = NULL;
 122 }
 123
 124 static inline void spu_add_wq(wait_queue_head_t * wq, wait_queue_t * wait,
 125                               int prio)
 126 {
 127         prepare_to_wait_exclusive(wq, wait, TASK_INTERRUPTIBLE);
 128         set_bit(prio, spu_prio->bitmap);
 129 }
 130
 131 static inline void spu_del_wq(wait_queue_head_t * wq, wait_queue_t * wait,
 132                               int prio)
 133 {
 134         u64 flags;
 135
 136         __set_current_state(TASK_RUNNING);
 137
 138         spin_lock_irqsave(&wq->lock, flags);
 139
 140         remove_wait_queue_locked(wq, wait);
 141         if (list_empty(&wq->task_list))
 142                 clear_bit(prio, spu_prio->bitmap);
 143
 144         spin_unlock_irqrestore(&wq->lock, flags);
 145 }
 146
 147 static void spu_prio_wait(struct spu_context *ctx, u64 flags)
 148 {
 149         int prio = current->prio;
 150         wait_queue_head_t *wq = &spu_prio->waitq[prio];
 151         DEFINE_WAIT(wait);
 152
 153         if (ctx->spu)
 154                 return;
 155
 156         spu_add_wq(wq, &wait, prio);
 157
 158         if (!signal_pending(current)) {
 159                 up_write(&ctx->state_sema);
 160                 pr_debug("%s: pid=%d prio=%d\n", __FUNCTION__,
 161                          current->pid, current->prio);
 162                 schedule();
 163                 down_write(&ctx->state_sema);
 164         }
 165
 166         spu_del_wq(wq, &wait, prio);
 167 }
 168
 169 static void spu_prio_wakeup(void)
 170 {
 171         int best = sched_find_first_bit(spu_prio->bitmap);
 172         if (best < MAX_PRIO) {
 173                 wait_queue_head_t *wq = &spu_prio->waitq[best];
 174                 wake_up_interruptible_nr(wq, 1);
 175         }
 176 }
 177
 178 static int get_active_spu(struct spu *spu)
 179 {
 180         int node = spu->node;
 181         struct spu *tmp;
 182         int rc = 0;
 183
 184         mutex_lock(&spu_prio->active_mutex[node]);
 185         list_for_each_entry(tmp, &spu_prio->active_list[node], list) {
 186                 if (tmp == spu) {
 187                         list_del_init(&spu->list);
 188                         rc = 1;
 189                         break;
 190                 }
 191         }
 192         mutex_unlock(&spu_prio->active_mutex[node]);
 193         return rc;
 194 }
 195
 196 static void put_active_spu(struct spu *spu)
 197 {
 198         int node = spu->node;
 199
 200         mutex_lock(&spu_prio->active_mutex[node]);
 201         list_add_tail(&spu->list, &spu_prio->active_list[node]);
 202         mutex_unlock(&spu_prio->active_mutex[node]);
 203 }
 204
 205 static struct spu *spu_get_idle(struct spu_context *ctx, u64 flags)
 206 {
 207         struct spu *spu = NULL;
 208         int node = cpu_to_node(raw_smp_processor_id());
 209         int n;
 210
 211         for (n = 0; n < MAX_NUMNODES; n++, node++) {
 212                 node = (node < MAX_NUMNODES) ? node : 0;
 213                 if (!node_allowed(node))
 214                         continue;
 215                 spu = spu_alloc_node(node);
 216                 if (spu)
 217                         break;
 218         }
 219         return spu;
 220 }
 221
 222 static inline struct spu *spu_get(struct spu_context *ctx, u64 flags)
 223 {
 224         /* Future: spu_get_idle() if possible,
 225          * otherwise try to preempt an active
 226          * context.
 227          */
 228         return spu_get_idle(ctx, flags);
 229 }
 230
 231 /* The three externally callable interfaces
 232  * for the scheduler begin here.
 233  *
 234  *      spu_activate    - bind a context to SPU, waiting as needed.
 235  *      spu_deactivate  - unbind a context from its SPU.
 236  *      spu_yield       - yield an SPU if others are waiting.
 237  */
 238
 239 int spu_activate(struct spu_context *ctx, u64 flags)
 240 {
 241         struct spu *spu;
 242         int ret = 0;
 243
 244         for (;;) {
 245                 if (ctx->spu)
 246                         return 0;
 247                 spu = spu_get(ctx, flags);
 248                 if (spu != NULL) {
 249                         if (ctx->spu != NULL) {
 250                                 spu_free(spu);
 251                                 spu_prio_wakeup();
 252                                 break;
 253                         }
 254                         bind_context(spu, ctx);
 255                         put_active_spu(spu);
 256                         break;
 257                 }
 258                 spu_prio_wait(ctx, flags);
 259                 if (signal_pending(current)) {
 260                         ret = -ERESTARTSYS;
 261                         spu_prio_wakeup();
 262                         break;
 263                 }
 264         }
 265         return ret;
 266 }
 267
 268 void spu_deactivate(struct spu_context *ctx)
 269 {
 270         struct spu *spu;
 271         int needs_idle;
 272
 273         spu = ctx->spu;
 274         if (!spu)
 275                 return;
 276         needs_idle = get_active_spu(spu);
 277         unbind_context(spu, ctx);
 278         if (needs_idle) {
 279                 spu_free(spu);
 280                 spu_prio_wakeup();
 281         }
 282 }
 283
 284 void spu_yield(struct spu_context *ctx)
 285 {
 286         struct spu *spu;
 287         int need_yield = 0;
 288
 289         if (down_write_trylock(&ctx->state_sema)) {
 290                 if ((spu = ctx->spu) != NULL) {
 291                         int best = sched_find_first_bit(spu_prio->bitmap);
 292                         if (best < MAX_PRIO) {
 293                                 pr_debug("%s: yielding SPU %d NODE %d\n",
 294                                          __FUNCTION__, spu->number, spu->node);
 295                                 spu_deactivate(ctx);
 296                                 ctx->state = SPU_STATE_SAVED;
 297                                 need_yield = 1;
 298                         } else {
 299                                 spu->prio = MAX_PRIO;
 300                         }
 301                 }
 302                 up_write(&ctx->state_sema);
 303         }
 304         if (unlikely(need_yield))
 305                 yield();
 306 }
 307
 308 int __init spu_sched_init(void)
 309 {
 310         int i;
 311
 312         spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
 313         if (!spu_prio) {
 314                 printk(KERN_WARNING "%s: Unable to allocate priority queue.\n",
 315                        __FUNCTION__);
 316                 return 1;
 317         }
 318         for (i = 0; i < MAX_PRIO; i++) {
 319                 init_waitqueue_head(&spu_prio->waitq[i]);
 320                 __clear_bit(i, spu_prio->bitmap);
 321         }
 322         __set_bit(MAX_PRIO, spu_prio->bitmap);
 323         for (i = 0; i < MAX_NUMNODES; i++) {
 324                 mutex_init(&spu_prio->active_mutex[i]);
 325                 INIT_LIST_HEAD(&spu_prio->active_list[i]);
 326         }
 327         return 0;
 328 }
 329
 330 void __exit spu_sched_exit(void)
 331 {
 332         struct spu *spu, *tmp;
 333         int node;
 334
 335         for (node = 0; node < MAX_NUMNODES; node++) {
 336                 mutex_lock(&spu_prio->active_mutex[node]);
 337                 list_for_each_entry_safe(spu, tmp, &spu_prio->active_list[node],
 338                                          list) {
 339                         list_del_init(&spu->list);
 340                         spu_free(spu);
 341                 }
 342                 mutex_unlock(&spu_prio->active_mutex[node]);
 343         }
 344         kfree(spu_prio);
 345 }