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/poll.h> 29 #include <linux/ptrace.h> 30 #include <linux/slab.h> 31 #include <linux/wait.h> 32 33 #include <asm/io.h> 34 #include <asm/prom.h> 35 #include <asm/semaphore.h> 36 #include <asm/spu.h> 37 #include <asm/mmu_context.h> 38 39 #include "interrupt.h" 40 41 static int __spu_trap_invalid_dma(struct spu *spu) 42 { 43 pr_debug("%s\n", __FUNCTION__); 44 force_sig(SIGBUS, /* info, */ current); 45 return 0; 46 } 47 48 static int __spu_trap_dma_align(struct spu *spu) 49 { 50 pr_debug("%s\n", __FUNCTION__); 51 force_sig(SIGBUS, /* info, */ current); 52 return 0; 53 } 54 55 static int __spu_trap_error(struct spu *spu) 56 { 57 pr_debug("%s\n", __FUNCTION__); 58 force_sig(SIGILL, /* info, */ current); 59 return 0; 60 } 61 62 static void spu_restart_dma(struct spu *spu) 63 { 64 struct spu_priv2 __iomem *priv2 = spu->priv2; 65 66 if (!test_bit(SPU_CONTEXT_SWITCH_PENDING, &spu->flags)) 67 out_be64(&priv2->mfc_control_RW, MFC_CNTL_RESTART_DMA_COMMAND); 68 } 69 70 static int __spu_trap_data_seg(struct spu *spu, unsigned long ea) 71 { 72 struct spu_priv2 __iomem *priv2 = spu->priv2; 73 struct mm_struct *mm = spu->mm; 74 u64 esid, vsid; 75 76 pr_debug("%s\n", __FUNCTION__); 77 78 if (test_bit(SPU_CONTEXT_SWITCH_ACTIVE, &spu->flags)) { 79 /* SLBs are pre-loaded for context switch, so 80 * we should never get here! 81 */ 82 printk("%s: invalid access during switch!\n", __func__); 83 return 1; 84 } 85 if (!mm || (REGION_ID(ea) != USER_REGION_ID)) { 86 /* Future: support kernel segments so that drivers 87 * can use SPUs. 88 */ 89 pr_debug("invalid region access at %016lx\n", ea); 90 return 1; 91 } 92 93 esid = (ea & ESID_MASK) | SLB_ESID_V; 94 vsid = (get_vsid(mm->context.id, ea) << SLB_VSID_SHIFT) | SLB_VSID_USER; 95 if (in_hugepage_area(mm->context, ea)) 96 vsid |= SLB_VSID_L; 97 98 out_be64(&priv2->slb_index_W, spu->slb_replace); 99 out_be64(&priv2->slb_vsid_RW, vsid); 100 out_be64(&priv2->slb_esid_RW, esid); 101 102 spu->slb_replace++; 103 if (spu->slb_replace >= 8) 104 spu->slb_replace = 0; 105 106 spu_restart_dma(spu); 107 108 return 0; 109 } 110 111 extern int hash_page(unsigned long ea, unsigned long access, unsigned long trap); //XXX 112 static int __spu_trap_data_map(struct spu *spu, unsigned long ea, u64 dsisr) 113 { 114 pr_debug("%s\n", __FUNCTION__); 115 116 /* Handle kernel space hash faults immediately. 117 User hash faults need to be deferred to process context. */ 118 if ((dsisr & MFC_DSISR_PTE_NOT_FOUND) 119 && REGION_ID(ea) != USER_REGION_ID 120 && hash_page(ea, _PAGE_PRESENT, 0x300) == 0) { 121 spu_restart_dma(spu); 122 return 0; 123 } 124 125 if (test_bit(SPU_CONTEXT_SWITCH_ACTIVE, &spu->flags)) { 126 printk("%s: invalid access during switch!\n", __func__); 127 return 1; 128 } 129 130 spu->dar = ea; 131 spu->dsisr = dsisr; 132 mb(); 133 if (spu->stop_callback) 134 spu->stop_callback(spu); 135 return 0; 136 } 137 138 static int __spu_trap_mailbox(struct spu *spu) 139 { 140 if (spu->ibox_callback) 141 spu->ibox_callback(spu); 142 143 /* atomically disable SPU mailbox interrupts */ 144 spin_lock(&spu->register_lock); 145 spu_int_mask_and(spu, 2, ~0x1); 146 spin_unlock(&spu->register_lock); 147 return 0; 148 } 149 150 static int __spu_trap_stop(struct spu *spu) 151 { 152 pr_debug("%s\n", __FUNCTION__); 153 spu->stop_code = in_be32(&spu->problem->spu_status_R); 154 if (spu->stop_callback) 155 spu->stop_callback(spu); 156 return 0; 157 } 158 159 static int __spu_trap_halt(struct spu *spu) 160 { 161 pr_debug("%s\n", __FUNCTION__); 162 spu->stop_code = in_be32(&spu->problem->spu_status_R); 163 if (spu->stop_callback) 164 spu->stop_callback(spu); 165 return 0; 166 } 167 168 static int __spu_trap_tag_group(struct spu *spu) 169 { 170 pr_debug("%s\n", __FUNCTION__); 171 /* wake_up(&spu->dma_wq); */ 172 return 0; 173 } 174 175 static int __spu_trap_spubox(struct spu *spu) 176 { 177 if (spu->wbox_callback) 178 spu->wbox_callback(spu); 179 180 /* atomically disable SPU mailbox interrupts */ 181 spin_lock(&spu->register_lock); 182 spu_int_mask_and(spu, 2, ~0x10); 183 spin_unlock(&spu->register_lock); 184 return 0; 185 } 186 187 static irqreturn_t 188 spu_irq_class_0(int irq, void *data, struct pt_regs *regs) 189 { 190 struct spu *spu; 191 192 spu = data; 193 spu->class_0_pending = 1; 194 if (spu->stop_callback) 195 spu->stop_callback(spu); 196 197 return IRQ_HANDLED; 198 } 199 200 int 201 spu_irq_class_0_bottom(struct spu *spu) 202 { 203 unsigned long stat, mask; 204 205 spu->class_0_pending = 0; 206 207 mask = spu_int_mask_get(spu, 0); 208 stat = spu_int_stat_get(spu, 0); 209 210 stat &= mask; 211 212 if (stat & 1) /* invalid MFC DMA */ 213 __spu_trap_invalid_dma(spu); 214 215 if (stat & 2) /* invalid DMA alignment */ 216 __spu_trap_dma_align(spu); 217 218 if (stat & 4) /* error on SPU */ 219 __spu_trap_error(spu); 220 221 spu_int_stat_clear(spu, 0, stat); 222 223 return (stat & 0x7) ? -EIO : 0; 224 } 225 EXPORT_SYMBOL_GPL(spu_irq_class_0_bottom); 226 227 static irqreturn_t 228 spu_irq_class_1(int irq, void *data, struct pt_regs *regs) 229 { 230 struct spu *spu; 231 unsigned long stat, mask, dar, dsisr; 232 233 spu = data; 234 235 /* atomically read & clear class1 status. */ 236 spin_lock(&spu->register_lock); 237 mask = spu_int_mask_get(spu, 1); 238 stat = spu_int_stat_get(spu, 1) & mask; 239 dar = spu_mfc_dar_get(spu); 240 dsisr = spu_mfc_dsisr_get(spu); 241 if (stat & 2) /* mapping fault */ 242 spu_mfc_dsisr_set(spu, 0ul); 243 spu_int_stat_clear(spu, 1, stat); 244 spin_unlock(&spu->register_lock); 245 246 if (stat & 1) /* segment fault */ 247 __spu_trap_data_seg(spu, dar); 248 249 if (stat & 2) { /* mapping fault */ 250 __spu_trap_data_map(spu, dar, dsisr); 251 } 252 253 if (stat & 4) /* ls compare & suspend on get */ 254 ; 255 256 if (stat & 8) /* ls compare & suspend on put */ 257 ; 258 259 return stat ? IRQ_HANDLED : IRQ_NONE; 260 } 261 EXPORT_SYMBOL_GPL(spu_irq_class_1_bottom); 262 263 static irqreturn_t 264 spu_irq_class_2(int irq, void *data, struct pt_regs *regs) 265 { 266 struct spu *spu; 267 unsigned long stat; 268 unsigned long mask; 269 270 spu = data; 271 stat = spu_int_stat_get(spu, 2); 272 mask = spu_int_mask_get(spu, 2); 273 274 pr_debug("class 2 interrupt %d, %lx, %lx\n", irq, stat, mask); 275 276 stat &= mask; 277 278 if (stat & 1) /* PPC core mailbox */ 279 __spu_trap_mailbox(spu); 280 281 if (stat & 2) /* SPU stop-and-signal */ 282 __spu_trap_stop(spu); 283 284 if (stat & 4) /* SPU halted */ 285 __spu_trap_halt(spu); 286 287 if (stat & 8) /* DMA tag group complete */ 288 __spu_trap_tag_group(spu); 289 290 if (stat & 0x10) /* SPU mailbox threshold */ 291 __spu_trap_spubox(spu); 292 293 spu_int_stat_clear(spu, 2, stat); 294 return stat ? IRQ_HANDLED : IRQ_NONE; 295 } 296 297 static int 298 spu_request_irqs(struct spu *spu) 299 { 300 int ret; 301 int irq_base; 302 303 irq_base = IIC_NODE_STRIDE * spu->node + IIC_SPE_OFFSET; 304 305 snprintf(spu->irq_c0, sizeof (spu->irq_c0), "spe%02d.0", spu->number); 306 ret = request_irq(irq_base + spu->isrc, 307 spu_irq_class_0, 0, spu->irq_c0, spu); 308 if (ret) 309 goto out; 310 311 snprintf(spu->irq_c1, sizeof (spu->irq_c1), "spe%02d.1", spu->number); 312 ret = request_irq(irq_base + IIC_CLASS_STRIDE + spu->isrc, 313 spu_irq_class_1, 0, spu->irq_c1, spu); 314 if (ret) 315 goto out1; 316 317 snprintf(spu->irq_c2, sizeof (spu->irq_c2), "spe%02d.2", spu->number); 318 ret = request_irq(irq_base + 2*IIC_CLASS_STRIDE + spu->isrc, 319 spu_irq_class_2, 0, spu->irq_c2, spu); 320 if (ret) 321 goto out2; 322 goto out; 323 324 out2: 325 free_irq(irq_base + IIC_CLASS_STRIDE + spu->isrc, spu); 326 out1: 327 free_irq(irq_base + spu->isrc, spu); 328 out: 329 return ret; 330 } 331 332 static void 333 spu_free_irqs(struct spu *spu) 334 { 335 int irq_base; 336 337 irq_base = IIC_NODE_STRIDE * spu->node + IIC_SPE_OFFSET; 338 339 free_irq(irq_base + spu->isrc, spu); 340 free_irq(irq_base + IIC_CLASS_STRIDE + spu->isrc, spu); 341 free_irq(irq_base + 2*IIC_CLASS_STRIDE + spu->isrc, spu); 342 } 343 344 static LIST_HEAD(spu_list); 345 static DECLARE_MUTEX(spu_mutex); 346 347 static void spu_init_channels(struct spu *spu) 348 { 349 static const struct { 350 unsigned channel; 351 unsigned count; 352 } zero_list[] = { 353 { 0x00, 1, }, { 0x01, 1, }, { 0x03, 1, }, { 0x04, 1, }, 354 { 0x18, 1, }, { 0x19, 1, }, { 0x1b, 1, }, { 0x1d, 1, }, 355 }, count_list[] = { 356 { 0x00, 0, }, { 0x03, 0, }, { 0x04, 0, }, { 0x15, 16, }, 357 { 0x17, 1, }, { 0x18, 0, }, { 0x19, 0, }, { 0x1b, 0, }, 358 { 0x1c, 1, }, { 0x1d, 0, }, { 0x1e, 1, }, 359 }; 360 struct spu_priv2 __iomem *priv2; 361 int i; 362 363 priv2 = spu->priv2; 364 365 /* initialize all channel data to zero */ 366 for (i = 0; i < ARRAY_SIZE(zero_list); i++) { 367 int count; 368 369 out_be64(&priv2->spu_chnlcntptr_RW, zero_list[i].channel); 370 for (count = 0; count < zero_list[i].count; count++) 371 out_be64(&priv2->spu_chnldata_RW, 0); 372 } 373 374 /* initialize channel counts to meaningful values */ 375 for (i = 0; i < ARRAY_SIZE(count_list); i++) { 376 out_be64(&priv2->spu_chnlcntptr_RW, count_list[i].channel); 377 out_be64(&priv2->spu_chnlcnt_RW, count_list[i].count); 378 } 379 } 380 381 struct spu *spu_alloc(void) 382 { 383 struct spu *spu; 384 385 down(&spu_mutex); 386 if (!list_empty(&spu_list)) { 387 spu = list_entry(spu_list.next, struct spu, list); 388 list_del_init(&spu->list); 389 pr_debug("Got SPU %x %d\n", spu->isrc, spu->number); 390 } else { 391 pr_debug("No SPU left\n"); 392 spu = NULL; 393 } 394 up(&spu_mutex); 395 396 if (spu) 397 spu_init_channels(spu); 398 399 return spu; 400 } 401 EXPORT_SYMBOL_GPL(spu_alloc); 402 403 void spu_free(struct spu *spu) 404 { 405 down(&spu_mutex); 406 list_add_tail(&spu->list, &spu_list); 407 up(&spu_mutex); 408 } 409 EXPORT_SYMBOL_GPL(spu_free); 410 411 static int spu_handle_mm_fault(struct spu *spu) 412 { 413 struct mm_struct *mm = spu->mm; 414 struct vm_area_struct *vma; 415 u64 ea, dsisr, is_write; 416 int ret; 417 418 ea = spu->dar; 419 dsisr = spu->dsisr; 420 #if 0 421 if (!IS_VALID_EA(ea)) { 422 return -EFAULT; 423 } 424 #endif /* XXX */ 425 if (mm == NULL) { 426 return -EFAULT; 427 } 428 if (mm->pgd == NULL) { 429 return -EFAULT; 430 } 431 432 down_read(&mm->mmap_sem); 433 vma = find_vma(mm, ea); 434 if (!vma) 435 goto bad_area; 436 if (vma->vm_start <= ea) 437 goto good_area; 438 if (!(vma->vm_flags & VM_GROWSDOWN)) 439 goto bad_area; 440 #if 0 441 if (expand_stack(vma, ea)) 442 goto bad_area; 443 #endif /* XXX */ 444 good_area: 445 is_write = dsisr & MFC_DSISR_ACCESS_PUT; 446 if (is_write) { 447 if (!(vma->vm_flags & VM_WRITE)) 448 goto bad_area; 449 } else { 450 if (dsisr & MFC_DSISR_ACCESS_DENIED) 451 goto bad_area; 452 if (!(vma->vm_flags & (VM_READ | VM_EXEC))) 453 goto bad_area; 454 } 455 ret = 0; 456 switch (handle_mm_fault(mm, vma, ea, is_write)) { 457 case VM_FAULT_MINOR: 458 current->min_flt++; 459 break; 460 case VM_FAULT_MAJOR: 461 current->maj_flt++; 462 break; 463 case VM_FAULT_SIGBUS: 464 ret = -EFAULT; 465 goto bad_area; 466 case VM_FAULT_OOM: 467 ret = -ENOMEM; 468 goto bad_area; 469 default: 470 BUG(); 471 } 472 up_read(&mm->mmap_sem); 473 return ret; 474 475 bad_area: 476 up_read(&mm->mmap_sem); 477 return -EFAULT; 478 } 479 480 int spu_irq_class_1_bottom(struct spu *spu) 481 { 482 u64 ea, dsisr, access, error = 0UL; 483 int ret = 0; 484 485 ea = spu->dar; 486 dsisr = spu->dsisr; 487 if (dsisr & MFC_DSISR_PTE_NOT_FOUND) { 488 access = (_PAGE_PRESENT | _PAGE_USER); 489 access |= (dsisr & MFC_DSISR_ACCESS_PUT) ? _PAGE_RW : 0UL; 490 if (hash_page(ea, access, 0x300) != 0) 491 error |= CLASS1_ENABLE_STORAGE_FAULT_INTR; 492 } 493 if ((error & CLASS1_ENABLE_STORAGE_FAULT_INTR) || 494 (dsisr & MFC_DSISR_ACCESS_DENIED)) { 495 if ((ret = spu_handle_mm_fault(spu)) != 0) 496 error |= CLASS1_ENABLE_STORAGE_FAULT_INTR; 497 else 498 error &= ~CLASS1_ENABLE_STORAGE_FAULT_INTR; 499 } 500 spu->dar = 0UL; 501 spu->dsisr = 0UL; 502 if (!error) { 503 spu_restart_dma(spu); 504 } else { 505 __spu_trap_invalid_dma(spu); 506 } 507 return ret; 508 } 509 510 void spu_irq_setaffinity(struct spu *spu, int cpu) 511 { 512 u64 target = iic_get_target_id(cpu); 513 u64 route = target << 48 | target << 32 | target << 16; 514 spu_int_route_set(spu, route); 515 } 516 EXPORT_SYMBOL_GPL(spu_irq_setaffinity); 517 518 static void __iomem * __init map_spe_prop(struct device_node *n, 519 const char *name) 520 { 521 struct address_prop { 522 unsigned long address; 523 unsigned int len; 524 } __attribute__((packed)) *prop; 525 526 void *p; 527 int proplen; 528 529 p = get_property(n, name, &proplen); 530 if (proplen != sizeof (struct address_prop)) 531 return NULL; 532 533 prop = p; 534 535 return ioremap(prop->address, prop->len); 536 } 537 538 static void spu_unmap(struct spu *spu) 539 { 540 iounmap(spu->priv2); 541 iounmap(spu->priv1); 542 iounmap(spu->problem); 543 iounmap((u8 __iomem *)spu->local_store); 544 } 545 546 static int __init spu_map_device(struct spu *spu, struct device_node *spe) 547 { 548 char *prop; 549 int ret; 550 551 ret = -ENODEV; 552 prop = get_property(spe, "isrc", NULL); 553 if (!prop) 554 goto out; 555 spu->isrc = *(unsigned int *)prop; 556 557 spu->name = get_property(spe, "name", NULL); 558 if (!spu->name) 559 goto out; 560 561 prop = get_property(spe, "local-store", NULL); 562 if (!prop) 563 goto out; 564 spu->local_store_phys = *(unsigned long *)prop; 565 566 /* we use local store as ram, not io memory */ 567 spu->local_store = (void __force *)map_spe_prop(spe, "local-store"); 568 if (!spu->local_store) 569 goto out; 570 571 spu->problem= map_spe_prop(spe, "problem"); 572 if (!spu->problem) 573 goto out_unmap; 574 575 spu->priv1= map_spe_prop(spe, "priv1"); 576 /* priv1 is not available on a hypervisor */ 577 578 spu->priv2= map_spe_prop(spe, "priv2"); 579 if (!spu->priv2) 580 goto out_unmap; 581 ret = 0; 582 goto out; 583 584 out_unmap: 585 spu_unmap(spu); 586 out: 587 return ret; 588 } 589 590 static int __init find_spu_node_id(struct device_node *spe) 591 { 592 unsigned int *id; 593 struct device_node *cpu; 594 595 cpu = spe->parent->parent; 596 id = (unsigned int *)get_property(cpu, "node-id", NULL); 597 598 return id ? *id : 0; 599 } 600 601 static int __init create_spu(struct device_node *spe) 602 { 603 struct spu *spu; 604 int ret; 605 static int number; 606 607 ret = -ENOMEM; 608 spu = kmalloc(sizeof (*spu), GFP_KERNEL); 609 if (!spu) 610 goto out; 611 612 ret = spu_map_device(spu, spe); 613 if (ret) 614 goto out_free; 615 616 spu->node = find_spu_node_id(spe); 617 spu->stop_code = 0; 618 spu->slb_replace = 0; 619 spu->mm = NULL; 620 spu->ctx = NULL; 621 spu->rq = NULL; 622 spu->pid = 0; 623 spu->class_0_pending = 0; 624 spu->flags = 0UL; 625 spu->dar = 0UL; 626 spu->dsisr = 0UL; 627 spin_lock_init(&spu->register_lock); 628 629 spu_mfc_sdr_set(spu, mfspr(SPRN_SDR1)); 630 spu_mfc_sr1_set(spu, 0x33); 631 632 spu->ibox_callback = NULL; 633 spu->wbox_callback = NULL; 634 spu->stop_callback = NULL; 635 636 down(&spu_mutex); 637 spu->number = number++; 638 ret = spu_request_irqs(spu); 639 if (ret) 640 goto out_unmap; 641 642 list_add(&spu->list, &spu_list); 643 up(&spu_mutex); 644 645 pr_debug(KERN_DEBUG "Using SPE %s %02x %p %p %p %p %d\n", 646 spu->name, spu->isrc, spu->local_store, 647 spu->problem, spu->priv1, spu->priv2, spu->number); 648 goto out; 649 650 out_unmap: 651 up(&spu_mutex); 652 spu_unmap(spu); 653 out_free: 654 kfree(spu); 655 out: 656 return ret; 657 } 658 659 static void destroy_spu(struct spu *spu) 660 { 661 list_del_init(&spu->list); 662 663 spu_free_irqs(spu); 664 spu_unmap(spu); 665 kfree(spu); 666 } 667 668 static void cleanup_spu_base(void) 669 { 670 struct spu *spu, *tmp; 671 down(&spu_mutex); 672 list_for_each_entry_safe(spu, tmp, &spu_list, list) 673 destroy_spu(spu); 674 up(&spu_mutex); 675 } 676 module_exit(cleanup_spu_base); 677 678 static int __init init_spu_base(void) 679 { 680 struct device_node *node; 681 int ret; 682 683 ret = -ENODEV; 684 for (node = of_find_node_by_type(NULL, "spe"); 685 node; node = of_find_node_by_type(node, "spe")) { 686 ret = create_spu(node); 687 if (ret) { 688 printk(KERN_WARNING "%s: Error initializing %s\n", 689 __FUNCTION__, node->name); 690 cleanup_spu_base(); 691 break; 692 } 693 } 694 /* in some old firmware versions, the spe is called 'spc', so we 695 look for that as well */ 696 for (node = of_find_node_by_type(NULL, "spc"); 697 node; node = of_find_node_by_type(node, "spc")) { 698 ret = create_spu(node); 699 if (ret) { 700 printk(KERN_WARNING "%s: Error initializing %s\n", 701 __FUNCTION__, node->name); 702 cleanup_spu_base(); 703 break; 704 } 705 } 706 return ret; 707 } 708 module_init(init_spu_base); 709 710 MODULE_LICENSE("GPL"); 711 MODULE_AUTHOR("Arnd Bergmann <arndb@de.ibm.com>"); 712