arch/sparc64/kernel/pci_sun4v.c

   1 /* pci_sun4v.c: SUN4V specific PCI controller support.
   2  *
   3  * Copyright (C) 2006 David S. Miller (davem@davemloft.net)
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/types.h>
   8 #include <linux/pci.h>
   9 #include <linux/init.h>
  10 #include <linux/slab.h>
  11 #include <linux/interrupt.h>
  12 #include <linux/percpu.h>
  13
  14 #include <asm/pbm.h>
  15 #include <asm/iommu.h>
  16 #include <asm/irq.h>
  17 #include <asm/upa.h>
  18 #include <asm/pstate.h>
  19 #include <asm/oplib.h>
  20 #include <asm/hypervisor.h>
  21
  22 #include "pci_impl.h"
  23 #include "iommu_common.h"
  24
  25 #include "pci_sun4v.h"
  26
  27 #define PGLIST_NENTS    (PAGE_SIZE / sizeof(u64))
  28
  29 struct pci_iommu_batch {
  30         struct pci_dev  *pdev;          /* Device mapping is for.       */
  31         unsigned long   prot;           /* IOMMU page protections       */
  32         unsigned long   entry;          /* Index into IOTSB.            */
  33         u64             *pglist;        /* List of physical pages       */
  34         unsigned long   npages;         /* Number of pages in list.     */
  35 };
  36
  37 static DEFINE_PER_CPU(struct pci_iommu_batch, pci_iommu_batch);
  38
  39 /* Interrupts must be disabled.  */
  40 static inline void pci_iommu_batch_start(struct pci_dev *pdev, unsigned long prot, unsigned long entry)
  41 {
  42         struct pci_iommu_batch *p = &__get_cpu_var(pci_iommu_batch);
  43
  44         p->pdev         = pdev;
  45         p->prot         = prot;
  46         p->entry        = entry;
  47         p->npages       = 0;
  48 }
  49
  50 /* Interrupts must be disabled.  */
  51 static long pci_iommu_batch_flush(struct pci_iommu_batch *p)
  52 {
  53         struct pcidev_cookie *pcp = p->pdev->sysdata;
  54         unsigned long devhandle = pcp->pbm->devhandle;
  55         unsigned long prot = p->prot;
  56         unsigned long entry = p->entry;
  57         u64 *pglist = p->pglist;
  58         unsigned long npages = p->npages;
  59
  60         while (npages != 0) {
  61                 long num;
  62
  63                 num = pci_sun4v_iommu_map(devhandle, HV_PCI_TSBID(0, entry),
  64                                           npages, prot, __pa(pglist));
  65                 if (unlikely(num < 0)) {
  66                         if (printk_ratelimit())
  67                                 printk("pci_iommu_batch_flush: IOMMU map of "
  68                                        "[%08lx:%08lx:%lx:%lx:%lx] failed with "
  69                                        "status %ld\n",
  70                                        devhandle, HV_PCI_TSBID(0, entry),
  71                                        npages, prot, __pa(pglist), num);
  72                         return -1;
  73                 }
  74
  75                 entry += num;
  76                 npages -= num;
  77                 pglist += num;
  78         }
  79
  80         p->entry = entry;
  81         p->npages = 0;
  82
  83         return 0;
  84 }
  85
  86 /* Interrupts must be disabled.  */
  87 static inline long pci_iommu_batch_add(u64 phys_page)
  88 {
  89         struct pci_iommu_batch *p = &__get_cpu_var(pci_iommu_batch);
  90
  91         BUG_ON(p->npages >= PGLIST_NENTS);
  92
  93         p->pglist[p->npages++] = phys_page;
  94         if (p->npages == PGLIST_NENTS)
  95                 return pci_iommu_batch_flush(p);
  96
  97         return 0;
  98 }
  99
 100 /* Interrupts must be disabled.  */
 101 static inline long pci_iommu_batch_end(void)
 102 {
 103         struct pci_iommu_batch *p = &__get_cpu_var(pci_iommu_batch);
 104
 105         BUG_ON(p->npages >= PGLIST_NENTS);
 106
 107         return pci_iommu_batch_flush(p);
 108 }
 109
 110 static long pci_arena_alloc(struct pci_iommu_arena *arena, unsigned long npages)
 111 {
 112         unsigned long n, i, start, end, limit;
 113         int pass;
 114
 115         limit = arena->limit;
 116         start = arena->hint;
 117         pass = 0;
 118
 119 again:
 120         n = find_next_zero_bit(arena->map, limit, start);
 121         end = n + npages;
 122         if (unlikely(end >= limit)) {
 123                 if (likely(pass < 1)) {
 124                         limit = start;
 125                         start = 0;
 126                         pass++;
 127                         goto again;
 128                 } else {
 129                         /* Scanned the whole thing, give up. */
 130                         return -1;
 131                 }
 132         }
 133
 134         for (i = n; i < end; i++) {
 135                 if (test_bit(i, arena->map)) {
 136                         start = i + 1;
 137                         goto again;
 138                 }
 139         }
 140
 141         for (i = n; i < end; i++)
 142                 __set_bit(i, arena->map);
 143
 144         arena->hint = end;
 145
 146         return n;
 147 }
 148
 149 static void pci_arena_free(struct pci_iommu_arena *arena, unsigned long base, unsigned long npages)
 150 {
 151         unsigned long i;
 152
 153         for (i = base; i < (base + npages); i++)
 154                 __clear_bit(i, arena->map);
 155 }
 156
 157 static void *pci_4v_alloc_consistent(struct pci_dev *pdev, size_t size, dma_addr_t *dma_addrp, gfp_t gfp)
 158 {
 159         struct pcidev_cookie *pcp;
 160         struct pci_iommu *iommu;
 161         unsigned long flags, order, first_page, npages, n;
 162         void *ret;
 163         long entry;
 164
 165         size = IO_PAGE_ALIGN(size);
 166         order = get_order(size);
 167         if (unlikely(order >= MAX_ORDER))
 168                 return NULL;
 169
 170         npages = size >> IO_PAGE_SHIFT;
 171
 172         first_page = __get_free_pages(gfp, order);
 173         if (unlikely(first_page == 0UL))
 174                 return NULL;
 175
 176         memset((char *)first_page, 0, PAGE_SIZE << order);
 177
 178         pcp = pdev->sysdata;
 179         iommu = pcp->pbm->iommu;
 180
 181         spin_lock_irqsave(&iommu->lock, flags);
 182         entry = pci_arena_alloc(&iommu->arena, npages);
 183         spin_unlock_irqrestore(&iommu->lock, flags);
 184
 185         if (unlikely(entry < 0L))
 186                 goto arena_alloc_fail;
 187
 188         *dma_addrp = (iommu->page_table_map_base +
 189                       (entry << IO_PAGE_SHIFT));
 190         ret = (void *) first_page;
 191         first_page = __pa(first_page);
 192
 193         local_irq_save(flags);
 194
 195         pci_iommu_batch_start(pdev,
 196                               (HV_PCI_MAP_ATTR_READ |
 197                                HV_PCI_MAP_ATTR_WRITE),
 198                               entry);
 199
 200         for (n = 0; n < npages; n++) {
 201                 long err = pci_iommu_batch_add(first_page + (n * PAGE_SIZE));
 202                 if (unlikely(err < 0L))
 203                         goto iommu_map_fail;
 204         }
 205
 206         if (unlikely(pci_iommu_batch_end() < 0L))
 207                 goto iommu_map_fail;
 208
 209         local_irq_restore(flags);
 210
 211         return ret;
 212
 213 iommu_map_fail:
 214         /* Interrupts are disabled.  */
 215         spin_lock(&iommu->lock);
 216         pci_arena_free(&iommu->arena, entry, npages);
 217         spin_unlock_irqrestore(&iommu->lock, flags);
 218
 219 arena_alloc_fail:
 220         free_pages(first_page, order);
 221         return NULL;
 222 }
 223
 224 static void pci_4v_free_consistent(struct pci_dev *pdev, size_t size, void *cpu, dma_addr_t dvma)
 225 {
 226         struct pcidev_cookie *pcp;
 227         struct pci_iommu *iommu;
 228         unsigned long flags, order, npages, entry;
 229         u32 devhandle;
 230
 231         npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT;
 232         pcp = pdev->sysdata;
 233         iommu = pcp->pbm->iommu;
 234         devhandle = pcp->pbm->devhandle;
 235         entry = ((dvma - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
 236
 237         spin_lock_irqsave(&iommu->lock, flags);
 238
 239         pci_arena_free(&iommu->arena, entry, npages);
 240
 241         do {
 242                 unsigned long num;
 243
 244                 num = pci_sun4v_iommu_demap(devhandle, HV_PCI_TSBID(0, entry),
 245                                             npages);
 246                 entry += num;
 247                 npages -= num;
 248         } while (npages != 0);
 249
 250         spin_unlock_irqrestore(&iommu->lock, flags);
 251
 252         order = get_order(size);
 253         if (order < 10)
 254                 free_pages((unsigned long)cpu, order);
 255 }
 256
 257 static dma_addr_t pci_4v_map_single(struct pci_dev *pdev, void *ptr, size_t sz, int direction)
 258 {
 259         struct pcidev_cookie *pcp;
 260         struct pci_iommu *iommu;
 261         unsigned long flags, npages, oaddr;
 262         unsigned long i, base_paddr;
 263         u32 bus_addr, ret;
 264         unsigned long prot;
 265         long entry;
 266
 267         pcp = pdev->sysdata;
 268         iommu = pcp->pbm->iommu;
 269
 270         if (unlikely(direction == PCI_DMA_NONE))
 271                 goto bad;
 272
 273         oaddr = (unsigned long)ptr;
 274         npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK);
 275         npages >>= IO_PAGE_SHIFT;
 276
 277         spin_lock_irqsave(&iommu->lock, flags);
 278         entry = pci_arena_alloc(&iommu->arena, npages);
 279         spin_unlock_irqrestore(&iommu->lock, flags);
 280
 281         if (unlikely(entry < 0L))
 282                 goto bad;
 283
 284         bus_addr = (iommu->page_table_map_base +
 285                     (entry << IO_PAGE_SHIFT));
 286         ret = bus_addr | (oaddr & ~IO_PAGE_MASK);
 287         base_paddr = __pa(oaddr & IO_PAGE_MASK);
 288         prot = HV_PCI_MAP_ATTR_READ;
 289         if (direction != PCI_DMA_TODEVICE)
 290                 prot |= HV_PCI_MAP_ATTR_WRITE;
 291
 292         local_irq_save(flags);
 293
 294         pci_iommu_batch_start(pdev, prot, entry);
 295
 296         for (i = 0; i < npages; i++, base_paddr += IO_PAGE_SIZE) {
 297                 long err = pci_iommu_batch_add(base_paddr);
 298                 if (unlikely(err < 0L))
 299                         goto iommu_map_fail;
 300         }
 301         if (unlikely(pci_iommu_batch_end() < 0L))
 302                 goto iommu_map_fail;
 303
 304         local_irq_restore(flags);
 305
 306         return ret;
 307
 308 bad:
 309         if (printk_ratelimit())
 310                 WARN_ON(1);
 311         return PCI_DMA_ERROR_CODE;
 312
 313 iommu_map_fail:
 314         /* Interrupts are disabled.  */
 315         spin_lock(&iommu->lock);
 316         pci_arena_free(&iommu->arena, entry, npages);
 317         spin_unlock_irqrestore(&iommu->lock, flags);
 318
 319         return PCI_DMA_ERROR_CODE;
 320 }
 321
 322 static void pci_4v_unmap_single(struct pci_dev *pdev, dma_addr_t bus_addr, size_t sz, int direction)
 323 {
 324         struct pcidev_cookie *pcp;
 325         struct pci_iommu *iommu;
 326         unsigned long flags, npages;
 327         long entry;
 328         u32 devhandle;
 329
 330         if (unlikely(direction == PCI_DMA_NONE)) {
 331                 if (printk_ratelimit())
 332                         WARN_ON(1);
 333                 return;
 334         }
 335
 336         pcp = pdev->sysdata;
 337         iommu = pcp->pbm->iommu;
 338         devhandle = pcp->pbm->devhandle;
 339
 340         npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK);
 341         npages >>= IO_PAGE_SHIFT;
 342         bus_addr &= IO_PAGE_MASK;
 343
 344         spin_lock_irqsave(&iommu->lock, flags);
 345
 346         entry = (bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT;
 347         pci_arena_free(&iommu->arena, entry, npages);
 348
 349         do {
 350                 unsigned long num;
 351
 352                 num = pci_sun4v_iommu_demap(devhandle, HV_PCI_TSBID(0, entry),
 353                                             npages);
 354                 entry += num;
 355                 npages -= num;
 356         } while (npages != 0);
 357
 358         spin_unlock_irqrestore(&iommu->lock, flags);
 359 }
 360
 361 #define SG_ENT_PHYS_ADDRESS(SG) \
 362         (__pa(page_address((SG)->page)) + (SG)->offset)
 363
 364 static inline long fill_sg(long entry, struct pci_dev *pdev,
 365                            struct scatterlist *sg,
 366                            int nused, int nelems, unsigned long prot)
 367 {
 368         struct scatterlist *dma_sg = sg;
 369         struct scatterlist *sg_end = sg + nelems;
 370         unsigned long flags;
 371         int i;
 372
 373         local_irq_save(flags);
 374
 375         pci_iommu_batch_start(pdev, prot, entry);
 376
 377         for (i = 0; i < nused; i++) {
 378                 unsigned long pteval = ~0UL;
 379                 u32 dma_npages;
 380
 381                 dma_npages = ((dma_sg->dma_address & (IO_PAGE_SIZE - 1UL)) +
 382                               dma_sg->dma_length +
 383                               ((IO_PAGE_SIZE - 1UL))) >> IO_PAGE_SHIFT;
 384                 do {
 385                         unsigned long offset;
 386                         signed int len;
 387
 388                         /* If we are here, we know we have at least one
 389                          * more page to map.  So walk forward until we
 390                          * hit a page crossing, and begin creating new
 391                          * mappings from that spot.
 392                          */
 393                         for (;;) {
 394                                 unsigned long tmp;
 395
 396                                 tmp = SG_ENT_PHYS_ADDRESS(sg);
 397                                 len = sg->length;
 398                                 if (((tmp ^ pteval) >> IO_PAGE_SHIFT) != 0UL) {
 399                                         pteval = tmp & IO_PAGE_MASK;
 400                                         offset = tmp & (IO_PAGE_SIZE - 1UL);
 401                                         break;
 402                                 }
 403                                 if (((tmp ^ (tmp + len - 1UL)) >> IO_PAGE_SHIFT) != 0UL) {
 404                                         pteval = (tmp + IO_PAGE_SIZE) & IO_PAGE_MASK;
 405                                         offset = 0UL;
 406                                         len -= (IO_PAGE_SIZE - (tmp & (IO_PAGE_SIZE - 1UL)));
 407                                         break;
 408                                 }
 409                                 sg++;
 410                         }
 411
 412                         pteval = (pteval & IOPTE_PAGE);
 413                         while (len > 0) {
 414                                 long err;
 415
 416                                 err = pci_iommu_batch_add(pteval);
 417                                 if (unlikely(err < 0L))
 418                                         goto iommu_map_failed;
 419
 420                                 pteval += IO_PAGE_SIZE;
 421                                 len -= (IO_PAGE_SIZE - offset);
 422                                 offset = 0;
 423                                 dma_npages--;
 424                         }
 425
 426                         pteval = (pteval & IOPTE_PAGE) + len;
 427                         sg++;
 428
 429                         /* Skip over any tail mappings we've fully mapped,
 430                          * adjusting pteval along the way.  Stop when we
 431                          * detect a page crossing event.
 432                          */
 433                         while (sg < sg_end &&
 434                                (pteval << (64 - IO_PAGE_SHIFT)) != 0UL &&
 435                                (pteval == SG_ENT_PHYS_ADDRESS(sg)) &&
 436                                ((pteval ^
 437                                  (SG_ENT_PHYS_ADDRESS(sg) + sg->length - 1UL)) >> IO_PAGE_SHIFT) == 0UL) {
 438                                 pteval += sg->length;
 439                                 sg++;
 440                         }
 441                         if ((pteval << (64 - IO_PAGE_SHIFT)) == 0UL)
 442                                 pteval = ~0UL;
 443                 } while (dma_npages != 0);
 444                 dma_sg++;
 445         }
 446
 447         if (unlikely(pci_iommu_batch_end() < 0L))
 448                 goto iommu_map_failed;
 449
 450         local_irq_restore(flags);
 451         return 0;
 452
 453 iommu_map_failed:
 454         local_irq_restore(flags);
 455         return -1L;
 456 }
 457
 458 static int pci_4v_map_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int direction)
 459 {
 460         struct pcidev_cookie *pcp;
 461         struct pci_iommu *iommu;
 462         unsigned long flags, npages, prot;
 463         u32 dma_base;
 464         struct scatterlist *sgtmp;
 465         long entry, err;
 466         int used;
 467
 468         /* Fast path single entry scatterlists. */
 469         if (nelems == 1) {
 470                 sglist->dma_address =
 471                         pci_4v_map_single(pdev,
 472                                           (page_address(sglist->page) + sglist->offset),
 473                                           sglist->length, direction);
 474                 if (unlikely(sglist->dma_address == PCI_DMA_ERROR_CODE))
 475                         return 0;
 476                 sglist->dma_length = sglist->length;
 477                 return 1;
 478         }
 479
 480         pcp = pdev->sysdata;
 481         iommu = pcp->pbm->iommu;
 482
 483         if (unlikely(direction == PCI_DMA_NONE))
 484                 goto bad;
 485
 486         /* Step 1: Prepare scatter list. */
 487         npages = prepare_sg(sglist, nelems);
 488
 489         /* Step 2: Allocate a cluster and context, if necessary. */
 490         spin_lock_irqsave(&iommu->lock, flags);
 491         entry = pci_arena_alloc(&iommu->arena, npages);
 492         spin_unlock_irqrestore(&iommu->lock, flags);
 493
 494         if (unlikely(entry < 0L))
 495                 goto bad;
 496
 497         dma_base = iommu->page_table_map_base +
 498                 (entry << IO_PAGE_SHIFT);
 499
 500         /* Step 3: Normalize DMA addresses. */
 501         used = nelems;
 502
 503         sgtmp = sglist;
 504         while (used && sgtmp->dma_length) {
 505                 sgtmp->dma_address += dma_base;
 506                 sgtmp++;
 507                 used--;
 508         }
 509         used = nelems - used;
 510
 511         /* Step 4: Create the mappings. */
 512         prot = HV_PCI_MAP_ATTR_READ;
 513         if (direction != PCI_DMA_TODEVICE)
 514                 prot |= HV_PCI_MAP_ATTR_WRITE;
 515
 516         err = fill_sg(entry, pdev, sglist, used, nelems, prot);
 517         if (unlikely(err < 0L))
 518                 goto iommu_map_failed;
 519
 520         return used;
 521
 522 bad:
 523         if (printk_ratelimit())
 524                 WARN_ON(1);
 525         return 0;
 526
 527 iommu_map_failed:
 528         spin_lock_irqsave(&iommu->lock, flags);
 529         pci_arena_free(&iommu->arena, entry, npages);
 530         spin_unlock_irqrestore(&iommu->lock, flags);
 531
 532         return 0;
 533 }
 534
 535 static void pci_4v_unmap_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int direction)
 536 {
 537         struct pcidev_cookie *pcp;
 538         struct pci_iommu *iommu;
 539         unsigned long flags, i, npages;
 540         long entry;
 541         u32 devhandle, bus_addr;
 542
 543         if (unlikely(direction == PCI_DMA_NONE)) {
 544                 if (printk_ratelimit())
 545                         WARN_ON(1);
 546         }
 547
 548         pcp = pdev->sysdata;
 549         iommu = pcp->pbm->iommu;
 550         devhandle = pcp->pbm->devhandle;
 551
 552         bus_addr = sglist->dma_address & IO_PAGE_MASK;
 553
 554         for (i = 1; i < nelems; i++)
 555                 if (sglist[i].dma_length == 0)
 556                         break;
 557         i--;
 558         npages = (IO_PAGE_ALIGN(sglist[i].dma_address + sglist[i].dma_length) -
 559                   bus_addr) >> IO_PAGE_SHIFT;
 560
 561         entry = ((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
 562
 563         spin_lock_irqsave(&iommu->lock, flags);
 564
 565         pci_arena_free(&iommu->arena, entry, npages);
 566
 567         do {
 568                 unsigned long num;
 569
 570                 num = pci_sun4v_iommu_demap(devhandle, HV_PCI_TSBID(0, entry),
 571                                             npages);
 572                 entry += num;
 573                 npages -= num;
 574         } while (npages != 0);
 575
 576         spin_unlock_irqrestore(&iommu->lock, flags);
 577 }
 578
 579 static void pci_4v_dma_sync_single_for_cpu(struct pci_dev *pdev, dma_addr_t bus_addr, size_t sz, int direction)
 580 {
 581         /* Nothing to do... */
 582 }
 583
 584 static void pci_4v_dma_sync_sg_for_cpu(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int direction)
 585 {
 586         /* Nothing to do... */
 587 }
 588
 589 struct pci_iommu_ops pci_sun4v_iommu_ops = {
 590         .alloc_consistent               = pci_4v_alloc_consistent,
 591         .free_consistent                = pci_4v_free_consistent,
 592         .map_single                     = pci_4v_map_single,
 593         .unmap_single                   = pci_4v_unmap_single,
 594         .map_sg                         = pci_4v_map_sg,
 595         .unmap_sg                       = pci_4v_unmap_sg,
 596         .dma_sync_single_for_cpu        = pci_4v_dma_sync_single_for_cpu,
 597         .dma_sync_sg_for_cpu            = pci_4v_dma_sync_sg_for_cpu,
 598 };
 599
 600 /* SUN4V PCI configuration space accessors. */
 601
 602 struct pdev_entry {
 603         struct pdev_entry       *next;
 604         u32                     devhandle;
 605         unsigned int            bus;
 606         unsigned int            device;
 607         unsigned int            func;
 608 };
 609
 610 #define PDEV_HTAB_SIZE  16
 611 #define PDEV_HTAB_MASK  (PDEV_HTAB_SIZE - 1)
 612 static struct pdev_entry *pdev_htab[PDEV_HTAB_SIZE];
 613
 614 static inline unsigned int pdev_hashfn(u32 devhandle, unsigned int bus, unsigned int device, unsigned int func)
 615 {
 616         unsigned int val;
 617
 618         val = (devhandle ^ (devhandle >> 4));
 619         val ^= bus;
 620         val ^= device;
 621         val ^= func;
 622
 623         return val & PDEV_HTAB_MASK;
 624 }
 625
 626 static int pdev_htab_add(u32 devhandle, unsigned int bus, unsigned int device, unsigned int func)
 627 {
 628         struct pdev_entry *p = kmalloc(sizeof(*p), GFP_KERNEL);
 629         struct pdev_entry **slot;
 630
 631         if (!p)
 632                 return -ENOMEM;
 633
 634         slot = &pdev_htab[pdev_hashfn(devhandle, bus, device, func)];
 635         p->next = *slot;
 636         *slot = p;
 637
 638         p->devhandle = devhandle;
 639         p->bus = bus;
 640         p->device = device;
 641         p->func = func;
 642
 643         return 0;
 644 }
 645
 646 /* Recursively descend into the OBP device tree, rooted at toplevel_node,
 647  * looking for a PCI device matching bus and devfn.
 648  */
 649 static int obp_find(struct linux_prom_pci_registers *pregs, int toplevel_node, unsigned int bus, unsigned int devfn)
 650 {
 651         toplevel_node = prom_getchild(toplevel_node);
 652
 653         while (toplevel_node != 0) {
 654                 int ret = obp_find(pregs, toplevel_node, bus, devfn);
 655
 656                 if (ret != 0)
 657                         return ret;
 658
 659                 ret = prom_getproperty(toplevel_node, "reg", (char *) pregs,
 660                                        sizeof(*pregs) * PROMREG_MAX);
 661                 if (ret == 0 || ret == -1)
 662                         goto next_sibling;
 663
 664                 if (((pregs[0].phys_hi >> 16) & 0xff) == bus &&
 665                     ((pregs[0].phys_hi >> 8) & 0xff) == devfn)
 666                         break;
 667
 668         next_sibling:
 669                 toplevel_node = prom_getsibling(toplevel_node);
 670         }
 671
 672         return toplevel_node;
 673 }
 674
 675 static int pdev_htab_populate(struct pci_pbm_info *pbm)
 676 {
 677         struct linux_prom_pci_registers pr[PROMREG_MAX];
 678         u32 devhandle = pbm->devhandle;
 679         unsigned int bus;
 680
 681         for (bus = pbm->pci_first_busno; bus <= pbm->pci_last_busno; bus++) {
 682                 unsigned int devfn;
 683
 684                 for (devfn = 0; devfn < 256; devfn++) {
 685                         unsigned int device = PCI_SLOT(devfn);
 686                         unsigned int func = PCI_FUNC(devfn);
 687
 688                         if (obp_find(pr, pbm->prom_node, bus, devfn)) {
 689                                 int err = pdev_htab_add(devhandle, bus,
 690                                                         device, func);
 691                                 if (err)
 692                                         return err;
 693                         }
 694                 }
 695         }
 696
 697         return 0;
 698 }
 699
 700 static struct pdev_entry *pdev_find(u32 devhandle, unsigned int bus, unsigned int device, unsigned int func)
 701 {
 702         struct pdev_entry *p;
 703
 704         p = pdev_htab[pdev_hashfn(devhandle, bus, device, func)];
 705         while (p) {
 706                 if (p->devhandle == devhandle &&
 707                     p->bus == bus &&
 708                     p->device == device &&
 709                     p->func == func)
 710                         break;
 711
 712                 p = p->next;
 713         }
 714
 715         return p;
 716 }
 717
 718 static inline int pci_sun4v_out_of_range(struct pci_pbm_info *pbm, unsigned int bus, unsigned int device, unsigned int func)
 719 {
 720         if (bus < pbm->pci_first_busno ||
 721             bus > pbm->pci_last_busno)
 722                 return 1;
 723         return pdev_find(pbm->devhandle, bus, device, func) == NULL;
 724 }
 725
 726 static int pci_sun4v_read_pci_cfg(struct pci_bus *bus_dev, unsigned int devfn,
 727                                   int where, int size, u32 *value)
 728 {
 729         struct pci_pbm_info *pbm = bus_dev->sysdata;
 730         u32 devhandle = pbm->devhandle;
 731         unsigned int bus = bus_dev->number;
 732         unsigned int device = PCI_SLOT(devfn);
 733         unsigned int func = PCI_FUNC(devfn);
 734         unsigned long ret;
 735
 736         if (pci_sun4v_out_of_range(pbm, bus, device, func)) {
 737                 ret = ~0UL;
 738         } else {
 739                 ret = pci_sun4v_config_get(devhandle,
 740                                 HV_PCI_DEVICE_BUILD(bus, device, func),
 741                                 where, size);
 742 #if 0
 743                 printk("rcfg: [%x:%x:%x:%d]=[%lx]\n",
 744                        devhandle, HV_PCI_DEVICE_BUILD(bus, device, func),
 745                        where, size, ret);
 746 #endif
 747         }
 748         switch (size) {
 749         case 1:
 750                 *value = ret & 0xff;
 751                 break;
 752         case 2:
 753                 *value = ret & 0xffff;
 754                 break;
 755         case 4:
 756                 *value = ret & 0xffffffff;
 757                 break;
 758         };
 759
 760
 761         return PCIBIOS_SUCCESSFUL;
 762 }
 763
 764 static int pci_sun4v_write_pci_cfg(struct pci_bus *bus_dev, unsigned int devfn,
 765                                    int where, int size, u32 value)
 766 {
 767         struct pci_pbm_info *pbm = bus_dev->sysdata;
 768         u32 devhandle = pbm->devhandle;
 769         unsigned int bus = bus_dev->number;
 770         unsigned int device = PCI_SLOT(devfn);
 771         unsigned int func = PCI_FUNC(devfn);
 772         unsigned long ret;
 773
 774         if (pci_sun4v_out_of_range(pbm, bus, device, func)) {
 775                 /* Do nothing. */
 776         } else {
 777                 ret = pci_sun4v_config_put(devhandle,
 778                                 HV_PCI_DEVICE_BUILD(bus, device, func),
 779                                 where, size, value);
 780 #if 0
 781                 printk("wcfg: [%x:%x:%x:%d] v[%x] == [%lx]\n",
 782                        devhandle, HV_PCI_DEVICE_BUILD(bus, device, func),
 783                        where, size, value, ret);
 784 #endif
 785         }
 786         return PCIBIOS_SUCCESSFUL;
 787 }
 788
 789 static struct pci_ops pci_sun4v_ops = {
 790         .read =         pci_sun4v_read_pci_cfg,
 791         .write =        pci_sun4v_write_pci_cfg,
 792 };
 793
 794
 795 static void pbm_scan_bus(struct pci_controller_info *p,
 796                          struct pci_pbm_info *pbm)
 797 {
 798         struct pcidev_cookie *cookie = kmalloc(sizeof(*cookie), GFP_KERNEL);
 799
 800         if (!cookie) {
 801                 prom_printf("%s: Critical allocation failure.\n", pbm->name);
 802                 prom_halt();
 803         }
 804
 805         /* All we care about is the PBM. */
 806         memset(cookie, 0, sizeof(*cookie));
 807         cookie->pbm = pbm;
 808
 809         pbm->pci_bus = pci_scan_bus(pbm->pci_first_busno, p->pci_ops, pbm);
 810 #if 0
 811         pci_fixup_host_bridge_self(pbm->pci_bus);
 812         pbm->pci_bus->self->sysdata = cookie;
 813 #endif
 814         pci_fill_in_pbm_cookies(pbm->pci_bus, pbm,
 815                                 pbm->prom_node);
 816         pci_record_assignments(pbm, pbm->pci_bus);
 817         pci_assign_unassigned(pbm, pbm->pci_bus);
 818         pci_fixup_irq(pbm, pbm->pci_bus);
 819         pci_determine_66mhz_disposition(pbm, pbm->pci_bus);
 820         pci_setup_busmastering(pbm, pbm->pci_bus);
 821 }
 822
 823 static void pci_sun4v_scan_bus(struct pci_controller_info *p)
 824 {
 825         if (p->pbm_A.prom_node) {
 826                 p->pbm_A.is_66mhz_capable =
 827                         prom_getbool(p->pbm_A.prom_node, "66mhz-capable");
 828
 829                 pbm_scan_bus(p, &p->pbm_A);
 830         }
 831         if (p->pbm_B.prom_node) {
 832                 p->pbm_B.is_66mhz_capable =
 833                         prom_getbool(p->pbm_B.prom_node, "66mhz-capable");
 834
 835                 pbm_scan_bus(p, &p->pbm_B);
 836         }
 837
 838         /* XXX register error interrupt handlers XXX */
 839 }
 840
 841 static unsigned int pci_sun4v_irq_build(struct pci_pbm_info *pbm,
 842                                         struct pci_dev *pdev,
 843                                         unsigned int devino)
 844 {
 845         u32 devhandle = pbm->devhandle;
 846         int pil;
 847
 848         pil = 5;
 849         if (pdev) {
 850                 switch ((pdev->class >> 16) & 0xff) {
 851                 case PCI_BASE_CLASS_STORAGE:
 852                         pil = 5;
 853                         break;
 854
 855                 case PCI_BASE_CLASS_NETWORK:
 856                         pil = 6;
 857                         break;
 858
 859                 case PCI_BASE_CLASS_DISPLAY:
 860                         pil = 9;
 861                         break;
 862
 863                 case PCI_BASE_CLASS_MULTIMEDIA:
 864                 case PCI_BASE_CLASS_MEMORY:
 865                 case PCI_BASE_CLASS_BRIDGE:
 866                 case PCI_BASE_CLASS_SERIAL:
 867                         pil = 10;
 868                         break;
 869
 870                 default:
 871                         pil = 5;
 872                         break;
 873                 };
 874         }
 875         BUG_ON(PIL_RESERVED(pil));
 876
 877         return sun4v_build_irq(devhandle, devino, pil, IBF_PCI);
 878 }
 879
 880 static void pci_sun4v_base_address_update(struct pci_dev *pdev, int resource)
 881 {
 882         struct pcidev_cookie *pcp = pdev->sysdata;
 883         struct pci_pbm_info *pbm = pcp->pbm;
 884         struct resource *res, *root;
 885         u32 reg;
 886         int where, size, is_64bit;
 887
 888         res = &pdev->resource[resource];
 889         if (resource < 6) {
 890                 where = PCI_BASE_ADDRESS_0 + (resource * 4);
 891         } else if (resource == PCI_ROM_RESOURCE) {
 892                 where = pdev->rom_base_reg;
 893         } else {
 894                 /* Somebody might have asked allocation of a non-standard resource */
 895                 return;
 896         }
 897
 898         /* XXX 64-bit MEM handling is not %100 correct... XXX */
 899         is_64bit = 0;
 900         if (res->flags & IORESOURCE_IO)
 901                 root = &pbm->io_space;
 902         else {
 903                 root = &pbm->mem_space;
 904                 if ((res->flags & PCI_BASE_ADDRESS_MEM_TYPE_MASK)
 905                     == PCI_BASE_ADDRESS_MEM_TYPE_64)
 906                         is_64bit = 1;
 907         }
 908
 909         size = res->end - res->start;
 910         pci_read_config_dword(pdev, where, &reg);
 911         reg = ((reg & size) |
 912                (((u32)(res->start - root->start)) & ~size));
 913         if (resource == PCI_ROM_RESOURCE) {
 914                 reg |= PCI_ROM_ADDRESS_ENABLE;
 915                 res->flags |= IORESOURCE_ROM_ENABLE;
 916         }
 917         pci_write_config_dword(pdev, where, reg);
 918
 919         /* This knows that the upper 32-bits of the address
 920          * must be zero.  Our PCI common layer enforces this.
 921          */
 922         if (is_64bit)
 923                 pci_write_config_dword(pdev, where + 4, 0);
 924 }
 925
 926 static void pci_sun4v_resource_adjust(struct pci_dev *pdev,
 927                                       struct resource *res,
 928                                       struct resource *root)
 929 {
 930         res->start += root->start;
 931         res->end += root->start;
 932 }
 933
 934 /* Use ranges property to determine where PCI MEM, I/O, and Config
 935  * space are for this PCI bus module.
 936  */
 937 static void pci_sun4v_determine_mem_io_space(struct pci_pbm_info *pbm)
 938 {
 939         int i, saw_mem, saw_io;
 940
 941         saw_mem = saw_io = 0;
 942         for (i = 0; i < pbm->num_pbm_ranges; i++) {
 943                 struct linux_prom_pci_ranges *pr = &pbm->pbm_ranges[i];
 944                 unsigned long a;
 945                 int type;
 946
 947                 type = (pr->child_phys_hi >> 24) & 0x3;
 948                 a = (((unsigned long)pr->parent_phys_hi << 32UL) |
 949                      ((unsigned long)pr->parent_phys_lo  <<  0UL));
 950
 951                 switch (type) {
 952                 case 1:
 953                         /* 16-bit IO space, 16MB */
 954                         pbm->io_space.start = a;
 955                         pbm->io_space.end = a + ((16UL*1024UL*1024UL) - 1UL);
 956                         pbm->io_space.flags = IORESOURCE_IO;
 957                         saw_io = 1;
 958                         break;
 959
 960                 case 2:
 961                         /* 32-bit MEM space, 2GB */
 962                         pbm->mem_space.start = a;
 963                         pbm->mem_space.end = a + (0x80000000UL - 1UL);
 964                         pbm->mem_space.flags = IORESOURCE_MEM;
 965                         saw_mem = 1;
 966                         break;
 967
 968                 case 3:
 969                         /* XXX 64-bit MEM handling XXX */
 970
 971                 default:
 972                         break;
 973                 };
 974         }
 975
 976         if (!saw_io || !saw_mem) {
 977                 prom_printf("%s: Fatal error, missing %s PBM range.\n",
 978                             pbm->name,
 979                             (!saw_io ? "IO" : "MEM"));
 980                 prom_halt();
 981         }
 982
 983         printk("%s: PCI IO[%lx] MEM[%lx]\n",
 984                pbm->name,
 985                pbm->io_space.start,
 986                pbm->mem_space.start);
 987 }
 988
 989 static void pbm_register_toplevel_resources(struct pci_controller_info *p,
 990                                             struct pci_pbm_info *pbm)
 991 {
 992         pbm->io_space.name = pbm->mem_space.name = pbm->name;
 993
 994         request_resource(&ioport_resource, &pbm->io_space);
 995         request_resource(&iomem_resource, &pbm->mem_space);
 996         pci_register_legacy_regions(&pbm->io_space,
 997                                     &pbm->mem_space);
 998 }
 999
1000 static unsigned long probe_existing_entries(struct pci_pbm_info *pbm,
1001                                             struct pci_iommu *iommu)
1002 {
1003         struct pci_iommu_arena *arena = &iommu->arena;
1004         unsigned long i, cnt = 0;
1005         u32 devhandle;
1006
1007         devhandle = pbm->devhandle;
1008         for (i = 0; i < arena->limit; i++) {
1009                 unsigned long ret, io_attrs, ra;
1010
1011                 ret = pci_sun4v_iommu_getmap(devhandle,
1012                                              HV_PCI_TSBID(0, i),
1013                                              &io_attrs, &ra);
1014                 if (ret == HV_EOK) {
1015                         cnt++;
1016                         __set_bit(i, arena->map);
1017                 }
1018         }
1019
1020         return cnt;
1021 }
1022
1023 static void pci_sun4v_iommu_init(struct pci_pbm_info *pbm)
1024 {
1025         struct pci_iommu *iommu = pbm->iommu;
1026         unsigned long num_tsb_entries, sz;
1027         u32 vdma[2], dma_mask, dma_offset;
1028         int err, tsbsize;
1029
1030         err = prom_getproperty(pbm->prom_node, "virtual-dma",
1031                                (char *)&vdma[0], sizeof(vdma));
1032         if (err == 0 || err == -1) {
1033                 /* No property, use default values. */
1034                 vdma[0] = 0x80000000;
1035                 vdma[1] = 0x80000000;
1036         }
1037
1038         dma_mask = vdma[0];
1039         switch (vdma[1]) {
1040                 case 0x20000000:
1041                         dma_mask |= 0x1fffffff;
1042                         tsbsize = 64;
1043                         break;
1044
1045                 case 0x40000000:
1046                         dma_mask |= 0x3fffffff;
1047                         tsbsize = 128;
1048                         break;
1049
1050                 case 0x80000000:
1051                         dma_mask |= 0x7fffffff;
1052                         tsbsize = 256;
1053                         break;
1054
1055                 default:
1056                         prom_printf("PCI-SUN4V: strange virtual-dma size.\n");
1057                         prom_halt();
1058         };
1059
1060         tsbsize *= (8 * 1024);
1061
1062         num_tsb_entries = tsbsize / sizeof(iopte_t);
1063
1064         dma_offset = vdma[0];
1065
1066         /* Setup initial software IOMMU state. */
1067         spin_lock_init(&iommu->lock);
1068         iommu->ctx_lowest_free = 1;
1069         iommu->page_table_map_base = dma_offset;
1070         iommu->dma_addr_mask = dma_mask;
1071
1072         /* Allocate and initialize the free area map.  */
1073         sz = num_tsb_entries / 8;
1074         sz = (sz + 7UL) & ~7UL;
1075         iommu->arena.map = kmalloc(sz, GFP_KERNEL);
1076         if (!iommu->arena.map) {
1077                 prom_printf("PCI_IOMMU: Error, kmalloc(arena.map) failed.\n");
1078                 prom_halt();
1079         }
1080         memset(iommu->arena.map, 0, sz);
1081         iommu->arena.limit = num_tsb_entries;
1082
1083         sz = probe_existing_entries(pbm, iommu);
1084
1085         printk("%s: TSB entries [%lu], existing mapings [%lu]\n",
1086                pbm->name, num_tsb_entries, sz);
1087 }
1088
1089 static void pci_sun4v_get_bus_range(struct pci_pbm_info *pbm)
1090 {
1091         unsigned int busrange[2];
1092         int prom_node = pbm->prom_node;
1093         int err;
1094
1095         err = prom_getproperty(prom_node, "bus-range",
1096                                (char *)&busrange[0],
1097                                sizeof(busrange));
1098         if (err == 0 || err == -1) {
1099                 prom_printf("%s: Fatal error, no bus-range.\n", pbm->name);
1100                 prom_halt();
1101         }
1102
1103         pbm->pci_first_busno = busrange[0];
1104         pbm->pci_last_busno = busrange[1];
1105
1106 }
1107
1108 static void pci_sun4v_pbm_init(struct pci_controller_info *p, int prom_node, u32 devhandle)
1109 {
1110         struct pci_pbm_info *pbm;
1111         int err, i;
1112
1113         if (devhandle & 0x40)
1114                 pbm = &p->pbm_B;
1115         else
1116                 pbm = &p->pbm_A;
1117
1118         pbm->parent = p;
1119         pbm->prom_node = prom_node;
1120         pbm->pci_first_slot = 1;
1121
1122         pbm->devhandle = devhandle;
1123
1124         sprintf(pbm->name, "SUN4V-PCI%d PBM%c",
1125                 p->index, (pbm == &p->pbm_A ? 'A' : 'B'));
1126
1127         printk("%s: devhandle[%x] prom_node[%x:%x]\n",
1128                pbm->name, pbm->devhandle,
1129                pbm->prom_node, prom_getchild(pbm->prom_node));
1130
1131         prom_getstring(prom_node, "name",
1132                        pbm->prom_name, sizeof(pbm->prom_name));
1133
1134         err = prom_getproperty(prom_node, "ranges",
1135                                (char *) pbm->pbm_ranges,
1136                                sizeof(pbm->pbm_ranges));
1137         if (err == 0 || err == -1) {
1138                 prom_printf("%s: Fatal error, no ranges property.\n",
1139                             pbm->name);
1140                 prom_halt();
1141         }
1142
1143         pbm->num_pbm_ranges =
1144                 (err / sizeof(struct linux_prom_pci_ranges));
1145
1146         /* Mask out the top 8 bits of the ranges, leaving the real
1147          * physical address.
1148          */
1149         for (i = 0; i < pbm->num_pbm_ranges; i++)
1150                 pbm->pbm_ranges[i].parent_phys_hi &= 0x0fffffff;
1151
1152         pci_sun4v_determine_mem_io_space(pbm);
1153         pbm_register_toplevel_resources(p, pbm);
1154
1155         err = prom_getproperty(prom_node, "interrupt-map",
1156                                (char *)pbm->pbm_intmap,
1157                                sizeof(pbm->pbm_intmap));
1158         if (err == 0 || err == -1) {
1159                 prom_printf("%s: Fatal error, no interrupt-map property.\n",
1160                             pbm->name);
1161                 prom_halt();
1162         }
1163
1164         pbm->num_pbm_intmap = (err / sizeof(struct linux_prom_pci_intmap));
1165         err = prom_getproperty(prom_node, "interrupt-map-mask",
1166                                (char *)&pbm->pbm_intmask,
1167                                sizeof(pbm->pbm_intmask));
1168         if (err == 0 || err == -1) {
1169                 prom_printf("%s: Fatal error, no interrupt-map-mask.\n",
1170                             pbm->name);
1171                 prom_halt();
1172         }
1173
1174         pci_sun4v_get_bus_range(pbm);
1175         pci_sun4v_iommu_init(pbm);
1176
1177         pdev_htab_populate(pbm);
1178 }
1179
1180 void sun4v_pci_init(int node, char *model_name)
1181 {
1182         struct pci_controller_info *p;
1183         struct pci_iommu *iommu;
1184         struct linux_prom64_registers regs;
1185         u32 devhandle;
1186         int i;
1187
1188         prom_getproperty(node, "reg", (char *)&regs, sizeof(regs));
1189         devhandle = (regs.phys_addr >> 32UL) & 0x0fffffff;
1190
1191         for (p = pci_controller_root; p; p = p->next) {
1192                 struct pci_pbm_info *pbm;
1193
1194                 if (p->pbm_A.prom_node && p->pbm_B.prom_node)
1195                         continue;
1196
1197                 pbm = (p->pbm_A.prom_node ?
1198                        &p->pbm_A :
1199                        &p->pbm_B);
1200
1201                 if (pbm->devhandle == (devhandle ^ 0x40)) {
1202                         pci_sun4v_pbm_init(p, node, devhandle);
1203                         return;
1204                 }
1205         }
1206
1207         for_each_possible_cpu(i) {
1208                 unsigned long page = get_zeroed_page(GFP_ATOMIC);
1209
1210                 if (!page)
1211                         goto fatal_memory_error;
1212
1213                 per_cpu(pci_iommu_batch, i).pglist = (u64 *) page;
1214         }
1215
1216         p = kmalloc(sizeof(struct pci_controller_info), GFP_ATOMIC);
1217         if (!p)
1218                 goto fatal_memory_error;
1219
1220         memset(p, 0, sizeof(*p));
1221
1222         iommu = kmalloc(sizeof(struct pci_iommu), GFP_ATOMIC);
1223         if (!iommu)
1224                 goto fatal_memory_error;
1225
1226         memset(iommu, 0, sizeof(*iommu));
1227         p->pbm_A.iommu = iommu;
1228
1229         iommu = kmalloc(sizeof(struct pci_iommu), GFP_ATOMIC);
1230         if (!iommu)
1231                 goto fatal_memory_error;
1232
1233         memset(iommu, 0, sizeof(*iommu));
1234         p->pbm_B.iommu = iommu;
1235
1236         p->next = pci_controller_root;
1237         pci_controller_root = p;
1238
1239         p->index = pci_num_controllers++;
1240         p->pbms_same_domain = 0;
1241
1242         p->scan_bus = pci_sun4v_scan_bus;
1243         p->irq_build = pci_sun4v_irq_build;
1244         p->base_address_update = pci_sun4v_base_address_update;
1245         p->resource_adjust = pci_sun4v_resource_adjust;
1246         p->pci_ops = &pci_sun4v_ops;
1247
1248         /* Like PSYCHO and SCHIZO we have a 2GB aligned area
1249          * for memory space.
1250          */
1251         pci_memspace_mask = 0x7fffffffUL;
1252
1253         pci_sun4v_pbm_init(p, node, devhandle);
1254         return;
1255
1256 fatal_memory_error:
1257         prom_printf("SUN4V_PCI: Fatal memory allocation error.\n");
1258         prom_halt();
1259 }