arch/powerpc/platforms/cell/spu_base.c

   1 /*
   2  * Low-level SPU handling
   3  *
   4  * (C) Copyright IBM Deutschland Entwicklung GmbH 2005
   5  *
   6  * Author: Arnd Bergmann <arndb@de.ibm.com>
   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/interrupt.h>
  26 #include <linux/list.h>
  27 #include <linux/module.h>
  28 #include <linux/pci.h>
  29 #include <linux/poll.h>
  30 #include <linux/ptrace.h>
  31 #include <linux/slab.h>
  32 #include <linux/wait.h>
  33
  34 #include <asm/firmware.h>
  35 #include <asm/io.h>
  36 #include <asm/prom.h>
  37 #include <linux/mutex.h>
  38 #include <asm/spu.h>
  39 #include <asm/spu_priv1.h>
  40 #include <asm/mmu_context.h>
  41
  42 #include "interrupt.h"
  43
  44 const struct spu_priv1_ops *spu_priv1_ops;
  45
  46 EXPORT_SYMBOL_GPL(spu_priv1_ops);
  47
  48 static int __spu_trap_invalid_dma(struct spu *spu)
  49 {
  50         pr_debug("%s\n", __FUNCTION__);
  51         spu->dma_callback(spu, SPE_EVENT_INVALID_DMA);
  52         return 0;
  53 }
  54
  55 static int __spu_trap_dma_align(struct spu *spu)
  56 {
  57         pr_debug("%s\n", __FUNCTION__);
  58         spu->dma_callback(spu, SPE_EVENT_DMA_ALIGNMENT);
  59         return 0;
  60 }
  61
  62 static int __spu_trap_error(struct spu *spu)
  63 {
  64         pr_debug("%s\n", __FUNCTION__);
  65         spu->dma_callback(spu, SPE_EVENT_SPE_ERROR);
  66         return 0;
  67 }
  68
  69 static void spu_restart_dma(struct spu *spu)
  70 {
  71         struct spu_priv2 __iomem *priv2 = spu->priv2;
  72
  73         if (!test_bit(SPU_CONTEXT_SWITCH_PENDING, &spu->flags))
  74                 out_be64(&priv2->mfc_control_RW, MFC_CNTL_RESTART_DMA_COMMAND);
  75 }
  76
  77 static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
  78 {
  79         struct spu_priv2 __iomem *priv2 = spu->priv2;
  80         struct mm_struct *mm = spu->mm;
  81         u64 esid, vsid, llp;
  82
  83         pr_debug("%s\n", __FUNCTION__);
  84
  85         if (test_bit(SPU_CONTEXT_SWITCH_ACTIVE, &spu->flags)) {
  86                 /* SLBs are pre-loaded for context switch, so
  87                  * we should never get here!
  88                  */
  89                 printk("%s: invalid access during switch!\n", __func__);
  90                 return 1;
  91         }
  92         if (!mm || (REGION_ID(ea) != USER_REGION_ID)) {
  93                 /* Future: support kernel segments so that drivers
  94                  * can use SPUs.
  95                  */
  96                 pr_debug("invalid region access at %016lx\n", ea);
  97                 return 1;
  98         }
  99
 100         esid = (ea & ESID_MASK) | SLB_ESID_V;
 101 #ifdef CONFIG_HUGETLB_PAGE
 102         if (in_hugepage_area(mm->context, ea))
 103                 llp = mmu_psize_defs[mmu_huge_psize].sllp;
 104         else
 105 #endif
 106                 llp = mmu_psize_defs[mmu_virtual_psize].sllp;
 107         vsid = (get_vsid(mm->context.id, ea) << SLB_VSID_SHIFT) |
 108                         SLB_VSID_USER | llp;
 109
 110         out_be64(&priv2->slb_index_W, spu->slb_replace);
 111         out_be64(&priv2->slb_vsid_RW, vsid);
 112         out_be64(&priv2->slb_esid_RW, esid);
 113
 114         spu->slb_replace++;
 115         if (spu->slb_replace >= 8)
 116                 spu->slb_replace = 0;
 117
 118         spu_restart_dma(spu);
 119
 120         return 0;
 121 }
 122
 123 extern int hash_page(unsigned long ea, unsigned long access, unsigned long trap); //XXX
 124 static int __spu_trap_data_map(struct spu *spu, unsigned long ea, u64 dsisr)
 125 {
 126         pr_debug("%s, %lx, %lx\n", __FUNCTION__, dsisr, ea);
 127
 128         /* Handle kernel space hash faults immediately.
 129            User hash faults need to be deferred to process context. */
 130         if ((dsisr & MFC_DSISR_PTE_NOT_FOUND)
 131             && REGION_ID(ea) != USER_REGION_ID
 132             && hash_page(ea, _PAGE_PRESENT, 0x300) == 0) {
 133                 spu_restart_dma(spu);
 134                 return 0;
 135         }
 136
 137         if (test_bit(SPU_CONTEXT_SWITCH_ACTIVE, &spu->flags)) {
 138                 printk("%s: invalid access during switch!\n", __func__);
 139                 return 1;
 140         }
 141
 142         spu->dar = ea;
 143         spu->dsisr = dsisr;
 144         mb();
 145         spu->stop_callback(spu);
 146         return 0;
 147 }
 148
 149 static irqreturn_t
 150 spu_irq_class_0(int irq, void *data)
 151 {
 152         struct spu *spu;
 153
 154         spu = data;
 155         spu->class_0_pending = 1;
 156         spu->stop_callback(spu);
 157
 158         return IRQ_HANDLED;
 159 }
 160
 161 int
 162 spu_irq_class_0_bottom(struct spu *spu)
 163 {
 164         unsigned long stat, mask;
 165
 166         spu->class_0_pending = 0;
 167
 168         mask = spu_int_mask_get(spu, 0);
 169         stat = spu_int_stat_get(spu, 0);
 170
 171         stat &= mask;
 172
 173         if (stat & 1) /* invalid DMA alignment */
 174                 __spu_trap_dma_align(spu);
 175
 176         if (stat & 2) /* invalid MFC DMA */
 177                 __spu_trap_invalid_dma(spu);
 178
 179         if (stat & 4) /* error on SPU */
 180                 __spu_trap_error(spu);
 181
 182         spu_int_stat_clear(spu, 0, stat);
 183
 184         return (stat & 0x7) ? -EIO : 0;
 185 }
 186 EXPORT_SYMBOL_GPL(spu_irq_class_0_bottom);
 187
 188 static irqreturn_t
 189 spu_irq_class_1(int irq, void *data)
 190 {
 191         struct spu *spu;
 192         unsigned long stat, mask, dar, dsisr;
 193
 194         spu = data;
 195
 196         /* atomically read & clear class1 status. */
 197         spin_lock(&spu->register_lock);
 198         mask  = spu_int_mask_get(spu, 1);
 199         stat  = spu_int_stat_get(spu, 1) & mask;
 200         dar   = spu_mfc_dar_get(spu);
 201         dsisr = spu_mfc_dsisr_get(spu);
 202         if (stat & 2) /* mapping fault */
 203                 spu_mfc_dsisr_set(spu, 0ul);
 204         spu_int_stat_clear(spu, 1, stat);
 205         spin_unlock(&spu->register_lock);
 206         pr_debug("%s: %lx %lx %lx %lx\n", __FUNCTION__, mask, stat,
 207                         dar, dsisr);
 208
 209         if (stat & 1) /* segment fault */
 210                 __spu_trap_data_seg(spu, dar);
 211
 212         if (stat & 2) { /* mapping fault */
 213                 __spu_trap_data_map(spu, dar, dsisr);
 214         }
 215
 216         if (stat & 4) /* ls compare & suspend on get */
 217                 ;
 218
 219         if (stat & 8) /* ls compare & suspend on put */
 220                 ;
 221
 222         return stat ? IRQ_HANDLED : IRQ_NONE;
 223 }
 224 EXPORT_SYMBOL_GPL(spu_irq_class_1_bottom);
 225
 226 static irqreturn_t
 227 spu_irq_class_2(int irq, void *data)
 228 {
 229         struct spu *spu;
 230         unsigned long stat;
 231         unsigned long mask;
 232
 233         spu = data;
 234         spin_lock(&spu->register_lock);
 235         stat = spu_int_stat_get(spu, 2);
 236         mask = spu_int_mask_get(spu, 2);
 237         /* ignore interrupts we're not waiting for */
 238         stat &= mask;
 239         /*
 240          * mailbox interrupts (0x1 and 0x10) are level triggered.
 241          * mask them now before acknowledging.
 242          */
 243         if (stat & 0x11)
 244                 spu_int_mask_and(spu, 2, ~(stat & 0x11));
 245         /* acknowledge all interrupts before the callbacks */
 246         spu_int_stat_clear(spu, 2, stat);
 247         spin_unlock(&spu->register_lock);
 248
 249         pr_debug("class 2 interrupt %d, %lx, %lx\n", irq, stat, mask);
 250
 251         if (stat & 1)  /* PPC core mailbox */
 252                 spu->ibox_callback(spu);
 253
 254         if (stat & 2) /* SPU stop-and-signal */
 255                 spu->stop_callback(spu);
 256
 257         if (stat & 4) /* SPU halted */
 258                 spu->stop_callback(spu);
 259
 260         if (stat & 8) /* DMA tag group complete */
 261                 spu->mfc_callback(spu);
 262
 263         if (stat & 0x10) /* SPU mailbox threshold */
 264                 spu->wbox_callback(spu);
 265
 266         return stat ? IRQ_HANDLED : IRQ_NONE;
 267 }
 268
 269 static int spu_request_irqs(struct spu *spu)
 270 {
 271         int ret = 0;
 272
 273         if (spu->irqs[0] != NO_IRQ) {
 274                 snprintf(spu->irq_c0, sizeof (spu->irq_c0), "spe%02d.0",
 275                          spu->number);
 276                 ret = request_irq(spu->irqs[0], spu_irq_class_0,
 277                                   IRQF_DISABLED,
 278                                   spu->irq_c0, spu);
 279                 if (ret)
 280                         goto bail0;
 281         }
 282         if (spu->irqs[1] != NO_IRQ) {
 283                 snprintf(spu->irq_c1, sizeof (spu->irq_c1), "spe%02d.1",
 284                          spu->number);
 285                 ret = request_irq(spu->irqs[1], spu_irq_class_1,
 286                                   IRQF_DISABLED,
 287                                   spu->irq_c1, spu);
 288                 if (ret)
 289                         goto bail1;
 290         }
 291         if (spu->irqs[2] != NO_IRQ) {
 292                 snprintf(spu->irq_c2, sizeof (spu->irq_c2), "spe%02d.2",
 293                          spu->number);
 294                 ret = request_irq(spu->irqs[2], spu_irq_class_2,
 295                                   IRQF_DISABLED,
 296                                   spu->irq_c2, spu);
 297                 if (ret)
 298                         goto bail2;
 299         }
 300         return 0;
 301
 302 bail2:
 303         if (spu->irqs[1] != NO_IRQ)
 304                 free_irq(spu->irqs[1], spu);
 305 bail1:
 306         if (spu->irqs[0] != NO_IRQ)
 307                 free_irq(spu->irqs[0], spu);
 308 bail0:
 309         return ret;
 310 }
 311
 312 static void spu_free_irqs(struct spu *spu)
 313 {
 314         if (spu->irqs[0] != NO_IRQ)
 315                 free_irq(spu->irqs[0], spu);
 316         if (spu->irqs[1] != NO_IRQ)
 317                 free_irq(spu->irqs[1], spu);
 318         if (spu->irqs[2] != NO_IRQ)
 319                 free_irq(spu->irqs[2], spu);
 320 }
 321
 322 static struct list_head spu_list[MAX_NUMNODES];
 323 static DEFINE_MUTEX(spu_mutex);
 324
 325 static void spu_init_channels(struct spu *spu)
 326 {
 327         static const struct {
 328                  unsigned channel;
 329                  unsigned count;
 330         } zero_list[] = {
 331                 { 0x00, 1, }, { 0x01, 1, }, { 0x03, 1, }, { 0x04, 1, },
 332                 { 0x18, 1, }, { 0x19, 1, }, { 0x1b, 1, }, { 0x1d, 1, },
 333         }, count_list[] = {
 334                 { 0x00, 0, }, { 0x03, 0, }, { 0x04, 0, }, { 0x15, 16, },
 335                 { 0x17, 1, }, { 0x18, 0, }, { 0x19, 0, }, { 0x1b, 0, },
 336                 { 0x1c, 1, }, { 0x1d, 0, }, { 0x1e, 1, },
 337         };
 338         struct spu_priv2 __iomem *priv2;
 339         int i;
 340
 341         priv2 = spu->priv2;
 342
 343         /* initialize all channel data to zero */
 344         for (i = 0; i < ARRAY_SIZE(zero_list); i++) {
 345                 int count;
 346
 347                 out_be64(&priv2->spu_chnlcntptr_RW, zero_list[i].channel);
 348                 for (count = 0; count < zero_list[i].count; count++)
 349                         out_be64(&priv2->spu_chnldata_RW, 0);
 350         }
 351
 352         /* initialize channel counts to meaningful values */
 353         for (i = 0; i < ARRAY_SIZE(count_list); i++) {
 354                 out_be64(&priv2->spu_chnlcntptr_RW, count_list[i].channel);
 355                 out_be64(&priv2->spu_chnlcnt_RW, count_list[i].count);
 356         }
 357 }
 358
 359 struct spu *spu_alloc_node(int node)
 360 {
 361         struct spu *spu = NULL;
 362
 363         mutex_lock(&spu_mutex);
 364         if (!list_empty(&spu_list[node])) {
 365                 spu = list_entry(spu_list[node].next, struct spu, list);
 366                 list_del_init(&spu->list);
 367                 pr_debug("Got SPU %x %d %d\n",
 368                          spu->isrc, spu->number, spu->node);
 369                 spu_init_channels(spu);
 370         }
 371         mutex_unlock(&spu_mutex);
 372
 373         return spu;
 374 }
 375 EXPORT_SYMBOL_GPL(spu_alloc_node);
 376
 377 struct spu *spu_alloc(void)
 378 {
 379         struct spu *spu = NULL;
 380         int node;
 381
 382         for (node = 0; node < MAX_NUMNODES; node++) {
 383                 spu = spu_alloc_node(node);
 384                 if (spu)
 385                         break;
 386         }
 387
 388         return spu;
 389 }
 390
 391 void spu_free(struct spu *spu)
 392 {
 393         mutex_lock(&spu_mutex);
 394         list_add_tail(&spu->list, &spu_list[spu->node]);
 395         mutex_unlock(&spu_mutex);
 396 }
 397 EXPORT_SYMBOL_GPL(spu_free);
 398
 399 static int spu_handle_mm_fault(struct spu *spu)
 400 {
 401         struct mm_struct *mm = spu->mm;
 402         struct vm_area_struct *vma;
 403         u64 ea, dsisr, is_write;
 404         int ret;
 405
 406         ea = spu->dar;
 407         dsisr = spu->dsisr;
 408 #if 0
 409         if (!IS_VALID_EA(ea)) {
 410                 return -EFAULT;
 411         }
 412 #endif /* XXX */
 413         if (mm == NULL) {
 414                 return -EFAULT;
 415         }
 416         if (mm->pgd == NULL) {
 417                 return -EFAULT;
 418         }
 419
 420         down_read(&mm->mmap_sem);
 421         vma = find_vma(mm, ea);
 422         if (!vma)
 423                 goto bad_area;
 424         if (vma->vm_start <= ea)
 425                 goto good_area;
 426         if (!(vma->vm_flags & VM_GROWSDOWN))
 427                 goto bad_area;
 428 #if 0
 429         if (expand_stack(vma, ea))
 430                 goto bad_area;
 431 #endif /* XXX */
 432 good_area:
 433         is_write = dsisr & MFC_DSISR_ACCESS_PUT;
 434         if (is_write) {
 435                 if (!(vma->vm_flags & VM_WRITE))
 436                         goto bad_area;
 437         } else {
 438                 if (dsisr & MFC_DSISR_ACCESS_DENIED)
 439                         goto bad_area;
 440                 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
 441                         goto bad_area;
 442         }
 443         ret = 0;
 444         switch (handle_mm_fault(mm, vma, ea, is_write)) {
 445         case VM_FAULT_MINOR:
 446                 current->min_flt++;
 447                 break;
 448         case VM_FAULT_MAJOR:
 449                 current->maj_flt++;
 450                 break;
 451         case VM_FAULT_SIGBUS:
 452                 ret = -EFAULT;
 453                 goto bad_area;
 454         case VM_FAULT_OOM:
 455                 ret = -ENOMEM;
 456                 goto bad_area;
 457         default:
 458                 BUG();
 459         }
 460         up_read(&mm->mmap_sem);
 461         return ret;
 462
 463 bad_area:
 464         up_read(&mm->mmap_sem);
 465         return -EFAULT;
 466 }
 467
 468 int spu_irq_class_1_bottom(struct spu *spu)
 469 {
 470         u64 ea, dsisr, access, error = 0UL;
 471         int ret = 0;
 472
 473         ea = spu->dar;
 474         dsisr = spu->dsisr;
 475         if (dsisr & (MFC_DSISR_PTE_NOT_FOUND | MFC_DSISR_ACCESS_DENIED)) {
 476                 u64 flags;
 477
 478                 access = (_PAGE_PRESENT | _PAGE_USER);
 479                 access |= (dsisr & MFC_DSISR_ACCESS_PUT) ? _PAGE_RW : 0UL;
 480                 local_irq_save(flags);
 481                 if (hash_page(ea, access, 0x300) != 0)
 482                         error |= CLASS1_ENABLE_STORAGE_FAULT_INTR;
 483                 local_irq_restore(flags);
 484         }
 485         if (error & CLASS1_ENABLE_STORAGE_FAULT_INTR) {
 486                 if ((ret = spu_handle_mm_fault(spu)) != 0)
 487                         error |= CLASS1_ENABLE_STORAGE_FAULT_INTR;
 488                 else
 489                         error &= ~CLASS1_ENABLE_STORAGE_FAULT_INTR;
 490         }
 491         spu->dar = 0UL;
 492         spu->dsisr = 0UL;
 493         if (!error) {
 494                 spu_restart_dma(spu);
 495         } else {
 496                 __spu_trap_invalid_dma(spu);
 497         }
 498         return ret;
 499 }
 500
 501 static int __init find_spu_node_id(struct device_node *spe)
 502 {
 503         const unsigned int *id;
 504         struct device_node *cpu;
 505         cpu = spe->parent->parent;
 506         id = get_property(cpu, "node-id", NULL);
 507         return id ? *id : 0;
 508 }
 509
 510 static int __init cell_spuprop_present(struct spu *spu, struct device_node *spe,
 511                 const char *prop)
 512 {
 513         static DEFINE_MUTEX(add_spumem_mutex);
 514
 515         const struct address_prop {
 516                 unsigned long address;
 517                 unsigned int len;
 518         } __attribute__((packed)) *p;
 519         int proplen;
 520
 521         unsigned long start_pfn, nr_pages;
 522         struct pglist_data *pgdata;
 523         struct zone *zone;
 524         int ret;
 525
 526         p = get_property(spe, prop, &proplen);
 527         WARN_ON(proplen != sizeof (*p));
 528
 529         start_pfn = p->address >> PAGE_SHIFT;
 530         nr_pages = ((unsigned long)p->len + PAGE_SIZE - 1) >> PAGE_SHIFT;
 531
 532         pgdata = NODE_DATA(spu->nid);
 533         zone = pgdata->node_zones;
 534
 535         /* XXX rethink locking here */
 536         mutex_lock(&add_spumem_mutex);
 537         ret = __add_pages(zone, start_pfn, nr_pages);
 538         mutex_unlock(&add_spumem_mutex);
 539
 540         return ret;
 541 }
 542
 543 static void __iomem * __init map_spe_prop(struct spu *spu,
 544                 struct device_node *n, const char *name)
 545 {
 546         const struct address_prop {
 547                 unsigned long address;
 548                 unsigned int len;
 549         } __attribute__((packed)) *prop;
 550
 551         const void *p;
 552         int proplen;
 553         void __iomem *ret = NULL;
 554         int err = 0;
 555
 556         p = get_property(n, name, &proplen);
 557         if (proplen != sizeof (struct address_prop))
 558                 return NULL;
 559
 560         prop = p;
 561
 562         err = cell_spuprop_present(spu, n, name);
 563         if (err && (err != -EEXIST))
 564                 goto out;
 565
 566         ret = ioremap(prop->address, prop->len);
 567
 568  out:
 569         return ret;
 570 }
 571
 572 static void spu_unmap(struct spu *spu)
 573 {
 574         iounmap(spu->priv2);
 575         iounmap(spu->priv1);
 576         iounmap(spu->problem);
 577         iounmap((__force u8 __iomem *)spu->local_store);
 578 }
 579
 580 /* This function shall be abstracted for HV platforms */
 581 static int __init spu_map_interrupts_old(struct spu *spu, struct device_node *np)
 582 {
 583         unsigned int isrc;
 584         const u32 *tmp;
 585
 586         /* Get the interrupt source unit from the device-tree */
 587         tmp = get_property(np, "isrc", NULL);
 588         if (!tmp)
 589                 return -ENODEV;
 590         isrc = tmp[0];
 591
 592         /* Add the node number */
 593         isrc |= spu->node << IIC_IRQ_NODE_SHIFT;
 594         spu->isrc = isrc;
 595
 596         /* Now map interrupts of all 3 classes */
 597         spu->irqs[0] = irq_create_mapping(NULL, IIC_IRQ_CLASS_0 | isrc);
 598         spu->irqs[1] = irq_create_mapping(NULL, IIC_IRQ_CLASS_1 | isrc);
 599         spu->irqs[2] = irq_create_mapping(NULL, IIC_IRQ_CLASS_2 | isrc);
 600
 601         /* Right now, we only fail if class 2 failed */
 602         return spu->irqs[2] == NO_IRQ ? -EINVAL : 0;
 603 }
 604
 605 static int __init spu_map_device_old(struct spu *spu, struct device_node *node)
 606 {
 607         const char *prop;
 608         int ret;
 609
 610         ret = -ENODEV;
 611         spu->name = get_property(node, "name", NULL);
 612         if (!spu->name)
 613                 goto out;
 614
 615         prop = get_property(node, "local-store", NULL);
 616         if (!prop)
 617                 goto out;
 618         spu->local_store_phys = *(unsigned long *)prop;
 619
 620         /* we use local store as ram, not io memory */
 621         spu->local_store = (void __force *)
 622                 map_spe_prop(spu, node, "local-store");
 623         if (!spu->local_store)
 624                 goto out;
 625
 626         prop = get_property(node, "problem", NULL);
 627         if (!prop)
 628                 goto out_unmap;
 629         spu->problem_phys = *(unsigned long *)prop;
 630
 631         spu->problem= map_spe_prop(spu, node, "problem");
 632         if (!spu->problem)
 633                 goto out_unmap;
 634
 635         spu->priv1= map_spe_prop(spu, node, "priv1");
 636         /* priv1 is not available on a hypervisor */
 637
 638         spu->priv2= map_spe_prop(spu, node, "priv2");
 639         if (!spu->priv2)
 640                 goto out_unmap;
 641         ret = 0;
 642         goto out;
 643
 644 out_unmap:
 645         spu_unmap(spu);
 646 out:
 647         return ret;
 648 }
 649
 650 static int __init spu_map_interrupts(struct spu *spu, struct device_node *np)
 651 {
 652         struct of_irq oirq;
 653         int ret;
 654         int i;
 655
 656         for (i=0; i < 3; i++) {
 657                 ret = of_irq_map_one(np, i, &oirq);
 658                 if (ret)
 659                         goto err;
 660
 661                 ret = -EINVAL;
 662                 spu->irqs[i] = irq_create_of_mapping(oirq.controller,
 663                                         oirq.specifier, oirq.size);
 664                 if (spu->irqs[i] == NO_IRQ)
 665                         goto err;
 666         }
 667         return 0;
 668
 669 err:
 670         pr_debug("failed to map irq %x for spu %s\n", *oirq.specifier, spu->name);
 671         for (; i >= 0; i--) {
 672                 if (spu->irqs[i] != NO_IRQ)
 673                         irq_dispose_mapping(spu->irqs[i]);
 674         }
 675         return ret;
 676 }
 677
 678 static int spu_map_resource(struct device_node *node, int nr,
 679                 void __iomem** virt, unsigned long *phys)
 680 {
 681         struct resource resource = { };
 682         int ret;
 683
 684         ret = of_address_to_resource(node, 0, &resource);
 685         if (ret)
 686                 goto out;
 687
 688         if (phys)
 689                 *phys = resource.start;
 690         *virt = ioremap(resource.start, resource.end - resource.start);
 691         if (!*virt)
 692                 ret = -EINVAL;
 693
 694 out:
 695         return ret;
 696 }
 697
 698 static int __init spu_map_device(struct spu *spu, struct device_node *node)
 699 {
 700         int ret = -ENODEV;
 701         spu->name = get_property(node, "name", NULL);
 702         if (!spu->name)
 703                 goto out;
 704
 705         ret = spu_map_resource(node, 0, (void __iomem**)&spu->local_store,
 706                                         &spu->local_store_phys);
 707         if (ret)
 708                 goto out;
 709         ret = spu_map_resource(node, 1, (void __iomem**)&spu->problem,
 710                                         &spu->problem_phys);
 711         if (ret)
 712                 goto out_unmap;
 713         ret = spu_map_resource(node, 2, (void __iomem**)&spu->priv2,
 714                                         NULL);
 715         if (ret)
 716                 goto out_unmap;
 717
 718         if (!firmware_has_feature(FW_FEATURE_LPAR))
 719                 ret = spu_map_resource(node, 3, (void __iomem**)&spu->priv1,
 720                                         NULL);
 721         if (ret)
 722                 goto out_unmap;
 723         return 0;
 724
 725 out_unmap:
 726         spu_unmap(spu);
 727 out:
 728         pr_debug("failed to map spe %s: %d\n", spu->name, ret);
 729         return ret;
 730 }
 731
 732 struct sysdev_class spu_sysdev_class = {
 733         set_kset_name("spu")
 734 };
 735
 736 static ssize_t spu_show_isrc(struct sys_device *sysdev, char *buf)
 737 {
 738         struct spu *spu = container_of(sysdev, struct spu, sysdev);
 739         return sprintf(buf, "%d\n", spu->isrc);
 740
 741 }
 742 static SYSDEV_ATTR(isrc, 0400, spu_show_isrc, NULL);
 743
 744 extern int attach_sysdev_to_node(struct sys_device *dev, int nid);
 745
 746 static int spu_create_sysdev(struct spu *spu)
 747 {
 748         int ret;
 749
 750         spu->sysdev.id = spu->number;
 751         spu->sysdev.cls = &spu_sysdev_class;
 752         ret = sysdev_register(&spu->sysdev);
 753         if (ret) {
 754                 printk(KERN_ERR "Can't register SPU %d with sysfs\n",
 755                                 spu->number);
 756                 return ret;
 757         }
 758
 759         if (spu->isrc != 0)
 760                 sysdev_create_file(&spu->sysdev, &attr_isrc);
 761         sysfs_add_device_to_node(&spu->sysdev, spu->nid);
 762
 763         return 0;
 764 }
 765
 766 static void spu_destroy_sysdev(struct spu *spu)
 767 {
 768         sysdev_remove_file(&spu->sysdev, &attr_isrc);
 769         sysfs_remove_device_from_node(&spu->sysdev, spu->nid);
 770         sysdev_unregister(&spu->sysdev);
 771 }
 772
 773 static int __init create_spu(struct device_node *spe)
 774 {
 775         struct spu *spu;
 776         int ret;
 777         static int number;
 778
 779         ret = -ENOMEM;
 780         spu = kzalloc(sizeof (*spu), GFP_KERNEL);
 781         if (!spu)
 782                 goto out;
 783
 784         spu->node = find_spu_node_id(spe);
 785         if (spu->node >= MAX_NUMNODES) {
 786                 printk(KERN_WARNING "SPE %s on node %d ignored,"
 787                        " node number too big\n", spe->full_name, spu->node);
 788                 printk(KERN_WARNING "Check if CONFIG_NUMA is enabled.\n");
 789                 return -ENODEV;
 790         }
 791         spu->nid = of_node_to_nid(spe);
 792         if (spu->nid == -1)
 793                 spu->nid = 0;
 794
 795         ret = spu_map_device(spu, spe);
 796         /* try old method */
 797         if (ret)
 798                 ret = spu_map_device_old(spu, spe);
 799         if (ret)
 800                 goto out_free;
 801
 802         ret = spu_map_interrupts(spu, spe);
 803         if (ret)
 804                 ret = spu_map_interrupts_old(spu, spe);
 805         if (ret)
 806                 goto out_unmap;
 807         spin_lock_init(&spu->register_lock);
 808         spu_mfc_sdr_set(spu, mfspr(SPRN_SDR1));
 809         spu_mfc_sr1_set(spu, 0x33);
 810         mutex_lock(&spu_mutex);
 811
 812         spu->number = number++;
 813         ret = spu_request_irqs(spu);
 814         if (ret)
 815                 goto out_unlock;
 816
 817         ret = spu_create_sysdev(spu);
 818         if (ret)
 819                 goto out_free_irqs;
 820
 821         list_add(&spu->list, &spu_list[spu->node]);
 822         mutex_unlock(&spu_mutex);
 823
 824         pr_debug(KERN_DEBUG "Using SPE %s %02x %p %p %p %p %d\n",
 825                 spu->name, spu->isrc, spu->local_store,
 826                 spu->problem, spu->priv1, spu->priv2, spu->number);
 827         goto out;
 828
 829 out_free_irqs:
 830         spu_free_irqs(spu);
 831 out_unlock:
 832         mutex_unlock(&spu_mutex);
 833 out_unmap:
 834         spu_unmap(spu);
 835 out_free:
 836         kfree(spu);
 837 out:
 838         return ret;
 839 }
 840
 841 static void destroy_spu(struct spu *spu)
 842 {
 843         list_del_init(&spu->list);
 844
 845         spu_destroy_sysdev(spu);
 846         spu_free_irqs(spu);
 847         spu_unmap(spu);
 848         kfree(spu);
 849 }
 850
 851 static void cleanup_spu_base(void)
 852 {
 853         struct spu *spu, *tmp;
 854         int node;
 855
 856         mutex_lock(&spu_mutex);
 857         for (node = 0; node < MAX_NUMNODES; node++) {
 858                 list_for_each_entry_safe(spu, tmp, &spu_list[node], list)
 859                         destroy_spu(spu);
 860         }
 861         mutex_unlock(&spu_mutex);
 862         sysdev_class_unregister(&spu_sysdev_class);
 863 }
 864 module_exit(cleanup_spu_base);
 865
 866 static int __init init_spu_base(void)
 867 {
 868         struct device_node *node;
 869         int i, ret;
 870
 871         /* create sysdev class for spus */
 872         ret = sysdev_class_register(&spu_sysdev_class);
 873         if (ret)
 874                 return ret;
 875
 876         for (i = 0; i < MAX_NUMNODES; i++)
 877                 INIT_LIST_HEAD(&spu_list[i]);
 878
 879         ret = -ENODEV;
 880         for (node = of_find_node_by_type(NULL, "spe");
 881                         node; node = of_find_node_by_type(node, "spe")) {
 882                 ret = create_spu(node);
 883                 if (ret) {
 884                         printk(KERN_WARNING "%s: Error initializing %s\n",
 885                                 __FUNCTION__, node->name);
 886                         cleanup_spu_base();
 887                         break;
 888                 }
 889         }
 890         return ret;
 891 }
 892 module_init(init_spu_base);
 893
 894 MODULE_LICENSE("GPL");
 895 MODULE_AUTHOR("Arnd Bergmann <arndb@de.ibm.com>");