git.oblomov.eu Git - linux-2.6/blob - arch/sparc64/kernel/pci_sun4v.c

   1 /* pci_sun4v.c: SUN4V specific PCI controller support.
   2  *
   3  * Copyright (C) 2006, 2007 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 #include <linux/irq.h>
  14 #include <linux/msi.h>
  15 #include <linux/log2.h>
  16
  17 #include <asm/iommu.h>
  18 #include <asm/irq.h>
  19 #include <asm/upa.h>
  20 #include <asm/pstate.h>
  21 #include <asm/oplib.h>
  22 #include <asm/hypervisor.h>
  23 #include <asm/prom.h>
  24
  25 #include "pci_impl.h"
  26 #include "iommu_common.h"
  27
  28 #include "pci_sun4v.h"
  29
  30 static unsigned long vpci_major = 1;
  31 static unsigned long vpci_minor = 1;
  32
  33 #define PGLIST_NENTS    (PAGE_SIZE / sizeof(u64))
  34
  35 struct iommu_batch {
  36         struct device   *dev;           /* Device mapping is for.       */
  37         unsigned long   prot;           /* IOMMU page protections       */
  38         unsigned long   entry;          /* Index into IOTSB.            */
  39         u64             *pglist;        /* List of physical pages       */
  40         unsigned long   npages;         /* Number of pages in list.     */
  41 };
  42
  43 static DEFINE_PER_CPU(struct iommu_batch, iommu_batch);
  44
  45 /* Interrupts must be disabled.  */
  46 static inline void iommu_batch_start(struct device *dev, unsigned long prot, unsigned long entry)
  47 {
  48         struct iommu_batch *p = &__get_cpu_var(iommu_batch);
  49
  50         p->dev          = dev;
  51         p->prot         = prot;
  52         p->entry        = entry;
  53         p->npages       = 0;
  54 }
  55
  56 /* Interrupts must be disabled.  */
  57 static long iommu_batch_flush(struct iommu_batch *p)
  58 {
  59         struct pci_pbm_info *pbm = p->dev->archdata.host_controller;
  60         unsigned long devhandle = pbm->devhandle;
  61         unsigned long prot = p->prot;
  62         unsigned long entry = p->entry;
  63         u64 *pglist = p->pglist;
  64         unsigned long npages = p->npages;
  65
  66         while (npages != 0) {
  67                 long num;
  68
  69                 num = pci_sun4v_iommu_map(devhandle, HV_PCI_TSBID(0, entry),
  70                                           npages, prot, __pa(pglist));
  71                 if (unlikely(num < 0)) {
  72                         if (printk_ratelimit())
  73                                 printk("iommu_batch_flush: IOMMU map of "
  74                                        "[%08lx:%08lx:%lx:%lx:%lx] failed with "
  75                                        "status %ld\n",
  76                                        devhandle, HV_PCI_TSBID(0, entry),
  77                                        npages, prot, __pa(pglist), num);
  78                         return -1;
  79                 }
  80
  81                 entry += num;
  82                 npages -= num;
  83                 pglist += num;
  84         }
  85
  86         p->entry = entry;
  87         p->npages = 0;
  88
  89         return 0;
  90 }
  91
  92 /* Interrupts must be disabled.  */
  93 static inline long iommu_batch_add(u64 phys_page)
  94 {
  95         struct iommu_batch *p = &__get_cpu_var(iommu_batch);
  96
  97         BUG_ON(p->npages >= PGLIST_NENTS);
  98
  99         p->pglist[p->npages++] = phys_page;
 100         if (p->npages == PGLIST_NENTS)
 101                 return iommu_batch_flush(p);
 102
 103         return 0;
 104 }
 105
 106 /* Interrupts must be disabled.  */
 107 static inline long iommu_batch_end(void)
 108 {
 109         struct iommu_batch *p = &__get_cpu_var(iommu_batch);
 110
 111         BUG_ON(p->npages >= PGLIST_NENTS);
 112
 113         return iommu_batch_flush(p);
 114 }
 115
 116 static long arena_alloc(struct iommu_arena *arena, unsigned long npages)
 117 {
 118         unsigned long n, i, start, end, limit;
 119         int pass;
 120
 121         limit = arena->limit;
 122         start = arena->hint;
 123         pass = 0;
 124
 125 again:
 126         n = find_next_zero_bit(arena->map, limit, start);
 127         end = n + npages;
 128         if (unlikely(end >= limit)) {
 129                 if (likely(pass < 1)) {
 130                         limit = start;
 131                         start = 0;
 132                         pass++;
 133                         goto again;
 134                 } else {
 135                         /* Scanned the whole thing, give up. */
 136                         return -1;
 137                 }
 138         }
 139
 140         for (i = n; i < end; i++) {
 141                 if (test_bit(i, arena->map)) {
 142                         start = i + 1;
 143                         goto again;
 144                 }
 145         }
 146
 147         for (i = n; i < end; i++)
 148                 __set_bit(i, arena->map);
 149
 150         arena->hint = end;
 151
 152         return n;
 153 }
 154
 155 static void arena_free(struct iommu_arena *arena, unsigned long base,
 156                        unsigned long npages)
 157 {
 158         unsigned long i;
 159
 160         for (i = base; i < (base + npages); i++)
 161                 __clear_bit(i, arena->map);
 162 }
 163
 164 static void *dma_4v_alloc_coherent(struct device *dev, size_t size,
 165                                    dma_addr_t *dma_addrp, gfp_t gfp)
 166 {
 167         struct iommu *iommu;
 168         unsigned long flags, order, first_page, npages, n;
 169         void *ret;
 170         long entry;
 171
 172         size = IO_PAGE_ALIGN(size);
 173         order = get_order(size);
 174         if (unlikely(order >= MAX_ORDER))
 175                 return NULL;
 176
 177         npages = size >> IO_PAGE_SHIFT;
 178
 179         first_page = __get_free_pages(gfp, order);
 180         if (unlikely(first_page == 0UL))
 181                 return NULL;
 182
 183         memset((char *)first_page, 0, PAGE_SIZE << order);
 184
 185         iommu = dev->archdata.iommu;
 186
 187         spin_lock_irqsave(&iommu->lock, flags);
 188         entry = arena_alloc(&iommu->arena, npages);
 189         spin_unlock_irqrestore(&iommu->lock, flags);
 190
 191         if (unlikely(entry < 0L))
 192                 goto arena_alloc_fail;
 193
 194         *dma_addrp = (iommu->page_table_map_base +
 195                       (entry << IO_PAGE_SHIFT));
 196         ret = (void *) first_page;
 197         first_page = __pa(first_page);
 198
 199         local_irq_save(flags);
 200
 201         iommu_batch_start(dev,
 202                           (HV_PCI_MAP_ATTR_READ |
 203                            HV_PCI_MAP_ATTR_WRITE),
 204                           entry);
 205
 206         for (n = 0; n < npages; n++) {
 207                 long err = iommu_batch_add(first_page + (n * PAGE_SIZE));
 208                 if (unlikely(err < 0L))
 209                         goto iommu_map_fail;
 210         }
 211
 212         if (unlikely(iommu_batch_end() < 0L))
 213                 goto iommu_map_fail;
 214
 215         local_irq_restore(flags);
 216
 217         return ret;
 218
 219 iommu_map_fail:
 220         /* Interrupts are disabled.  */
 221         spin_lock(&iommu->lock);
 222         arena_free(&iommu->arena, entry, npages);
 223         spin_unlock_irqrestore(&iommu->lock, flags);
 224
 225 arena_alloc_fail:
 226         free_pages(first_page, order);
 227         return NULL;
 228 }
 229
 230 static void dma_4v_free_coherent(struct device *dev, size_t size, void *cpu,
 231                                  dma_addr_t dvma)
 232 {
 233         struct pci_pbm_info *pbm;
 234         struct iommu *iommu;
 235         unsigned long flags, order, npages, entry;
 236         u32 devhandle;
 237
 238         npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT;
 239         iommu = dev->archdata.iommu;
 240         pbm = dev->archdata.host_controller;
 241         devhandle = pbm->devhandle;
 242         entry = ((dvma - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
 243
 244         spin_lock_irqsave(&iommu->lock, flags);
 245
 246         arena_free(&iommu->arena, entry, npages);
 247
 248         do {
 249                 unsigned long num;
 250
 251                 num = pci_sun4v_iommu_demap(devhandle, HV_PCI_TSBID(0, entry),
 252                                             npages);
 253                 entry += num;
 254                 npages -= num;
 255         } while (npages != 0);
 256
 257         spin_unlock_irqrestore(&iommu->lock, flags);
 258
 259         order = get_order(size);
 260         if (order < 10)
 261                 free_pages((unsigned long)cpu, order);
 262 }
 263
 264 static dma_addr_t dma_4v_map_single(struct device *dev, void *ptr, size_t sz,
 265                                     enum dma_data_direction direction)
 266 {
 267         struct iommu *iommu;
 268         unsigned long flags, npages, oaddr;
 269         unsigned long i, base_paddr;
 270         u32 bus_addr, ret;
 271         unsigned long prot;
 272         long entry;
 273
 274         iommu = dev->archdata.iommu;
 275
 276         if (unlikely(direction == DMA_NONE))
 277                 goto bad;
 278
 279         oaddr = (unsigned long)ptr;
 280         npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK);
 281         npages >>= IO_PAGE_SHIFT;
 282
 283         spin_lock_irqsave(&iommu->lock, flags);
 284         entry = arena_alloc(&iommu->arena, npages);
 285         spin_unlock_irqrestore(&iommu->lock, flags);
 286
 287         if (unlikely(entry < 0L))
 288                 goto bad;
 289
 290         bus_addr = (iommu->page_table_map_base +
 291                     (entry << IO_PAGE_SHIFT));
 292         ret = bus_addr | (oaddr & ~IO_PAGE_MASK);
 293         base_paddr = __pa(oaddr & IO_PAGE_MASK);
 294         prot = HV_PCI_MAP_ATTR_READ;
 295         if (direction != DMA_TO_DEVICE)
 296                 prot |= HV_PCI_MAP_ATTR_WRITE;
 297
 298         local_irq_save(flags);
 299
 300         iommu_batch_start(dev, prot, entry);
 301
 302         for (i = 0; i < npages; i++, base_paddr += IO_PAGE_SIZE) {
 303                 long err = iommu_batch_add(base_paddr);
 304                 if (unlikely(err < 0L))
 305                         goto iommu_map_fail;
 306         }
 307         if (unlikely(iommu_batch_end() < 0L))
 308                 goto iommu_map_fail;
 309
 310         local_irq_restore(flags);
 311
 312         return ret;
 313
 314 bad:
 315         if (printk_ratelimit())
 316                 WARN_ON(1);
 317         return DMA_ERROR_CODE;
 318
 319 iommu_map_fail:
 320         /* Interrupts are disabled.  */
 321         spin_lock(&iommu->lock);
 322         arena_free(&iommu->arena, entry, npages);
 323         spin_unlock_irqrestore(&iommu->lock, flags);
 324
 325         return DMA_ERROR_CODE;
 326 }
 327
 328 static void dma_4v_unmap_single(struct device *dev, dma_addr_t bus_addr,
 329                                 size_t sz, enum dma_data_direction direction)
 330 {
 331         struct pci_pbm_info *pbm;
 332         struct iommu *iommu;
 333         unsigned long flags, npages;
 334         long entry;
 335         u32 devhandle;
 336
 337         if (unlikely(direction == DMA_NONE)) {
 338                 if (printk_ratelimit())
 339                         WARN_ON(1);
 340                 return;
 341         }
 342
 343         iommu = dev->archdata.iommu;
 344         pbm = dev->archdata.host_controller;
 345         devhandle = pbm->devhandle;
 346
 347         npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK);
 348         npages >>= IO_PAGE_SHIFT;
 349         bus_addr &= IO_PAGE_MASK;
 350
 351         spin_lock_irqsave(&iommu->lock, flags);
 352
 353         entry = (bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT;
 354         arena_free(&iommu->arena, entry, npages);
 355
 356         do {
 357                 unsigned long num;
 358
 359                 num = pci_sun4v_iommu_demap(devhandle, HV_PCI_TSBID(0, entry),
 360                                             npages);
 361                 entry += num;
 362                 npages -= num;
 363         } while (npages != 0);
 364
 365         spin_unlock_irqrestore(&iommu->lock, flags);
 366 }
 367
 368 #define SG_ENT_PHYS_ADDRESS(SG) \
 369         (__pa(page_address((SG)->page)) + (SG)->offset)
 370
 371 static inline long fill_sg(long entry, struct device *dev,
 372                            struct scatterlist *sg,
 373                            int nused, int nelems, unsigned long prot)
 374 {
 375         struct scatterlist *dma_sg = sg;
 376         struct scatterlist *sg_end = sg_last(sg, nelems);
 377         unsigned long flags;
 378         int i;
 379
 380         local_irq_save(flags);
 381
 382         iommu_batch_start(dev, prot, entry);
 383
 384         for (i = 0; i < nused; i++) {
 385                 unsigned long pteval = ~0UL;
 386                 u32 dma_npages;
 387
 388                 dma_npages = ((dma_sg->dma_address & (IO_PAGE_SIZE - 1UL)) +
 389                               dma_sg->dma_length +
 390                               ((IO_PAGE_SIZE - 1UL))) >> IO_PAGE_SHIFT;
 391                 do {
 392                         unsigned long offset;
 393                         signed int len;
 394
 395                         /* If we are here, we know we have at least one
 396                          * more page to map.  So walk forward until we
 397                          * hit a page crossing, and begin creating new
 398                          * mappings from that spot.
 399                          */
 400                         for (;;) {
 401                                 unsigned long tmp;
 402
 403                                 tmp = SG_ENT_PHYS_ADDRESS(sg);
 404                                 len = sg->length;
 405                                 if (((tmp ^ pteval) >> IO_PAGE_SHIFT) != 0UL) {
 406                                         pteval = tmp & IO_PAGE_MASK;
 407                                         offset = tmp & (IO_PAGE_SIZE - 1UL);
 408                                         break;
 409                                 }
 410                                 if (((tmp ^ (tmp + len - 1UL)) >> IO_PAGE_SHIFT) != 0UL) {
 411                                         pteval = (tmp + IO_PAGE_SIZE) & IO_PAGE_MASK;
 412                                         offset = 0UL;
 413                                         len -= (IO_PAGE_SIZE - (tmp & (IO_PAGE_SIZE - 1UL)));
 414                                         break;
 415                                 }
 416                                 sg = sg_next(sg);
 417                         }
 418
 419                         pteval = (pteval & IOPTE_PAGE);
 420                         while (len > 0) {
 421                                 long err;
 422
 423                                 err = iommu_batch_add(pteval);
 424                                 if (unlikely(err < 0L))
 425                                         goto iommu_map_failed;
 426
 427                                 pteval += IO_PAGE_SIZE;
 428                                 len -= (IO_PAGE_SIZE - offset);
 429                                 offset = 0;
 430                                 dma_npages--;
 431                         }
 432
 433                         pteval = (pteval & IOPTE_PAGE) + len;
 434                         sg = sg_next(sg);
 435
 436                         /* Skip over any tail mappings we've fully mapped,
 437                          * adjusting pteval along the way.  Stop when we
 438                          * detect a page crossing event.
 439                          */
 440                         while ((pteval << (64 - IO_PAGE_SHIFT)) != 0UL &&
 441                                (pteval == SG_ENT_PHYS_ADDRESS(sg)) &&
 442                                ((pteval ^
 443                                  (SG_ENT_PHYS_ADDRESS(sg) + sg->length - 1UL)) >> IO_PAGE_SHIFT) == 0UL) {
 444                                 pteval += sg->length;
 445                                 if (sg == sg_end)
 446                                         break;
 447                                 sg = sg_next(sg);
 448                         }
 449                         if ((pteval << (64 - IO_PAGE_SHIFT)) == 0UL)
 450                                 pteval = ~0UL;
 451                 } while (dma_npages != 0);
 452                 dma_sg = sg_next(dma_sg);
 453         }
 454
 455         if (unlikely(iommu_batch_end() < 0L))
 456                 goto iommu_map_failed;
 457
 458         local_irq_restore(flags);
 459         return 0;
 460
 461 iommu_map_failed:
 462         local_irq_restore(flags);
 463         return -1L;
 464 }
 465
 466 static int dma_4v_map_sg(struct device *dev, struct scatterlist *sglist,
 467                          int nelems, enum dma_data_direction direction)
 468 {
 469         struct iommu *iommu;
 470         unsigned long flags, npages, prot;
 471         u32 dma_base;
 472         struct scatterlist *sgtmp;
 473         long entry, err;
 474         int used;
 475
 476         /* Fast path single entry scatterlists. */
 477         if (nelems == 1) {
 478                 sglist->dma_address =
 479                         dma_4v_map_single(dev,
 480                                           (page_address(sglist->page) +
 481                                            sglist->offset),
 482                                           sglist->length, direction);
 483                 if (unlikely(sglist->dma_address == DMA_ERROR_CODE))
 484                         return 0;
 485                 sglist->dma_length = sglist->length;
 486                 return 1;
 487         }
 488
 489         iommu = dev->archdata.iommu;
 490
 491         if (unlikely(direction == DMA_NONE))
 492                 goto bad;
 493
 494         /* Step 1: Prepare scatter list. */
 495         npages = prepare_sg(sglist, nelems);
 496
 497         /* Step 2: Allocate a cluster and context, if necessary. */
 498         spin_lock_irqsave(&iommu->lock, flags);
 499         entry = arena_alloc(&iommu->arena, npages);
 500         spin_unlock_irqrestore(&iommu->lock, flags);
 501
 502         if (unlikely(entry < 0L))
 503                 goto bad;
 504
 505         dma_base = iommu->page_table_map_base +
 506                 (entry << IO_PAGE_SHIFT);
 507
 508         /* Step 3: Normalize DMA addresses. */
 509         used = nelems;
 510
 511         sgtmp = sglist;
 512         while (used && sgtmp->dma_length) {
 513                 sgtmp->dma_address += dma_base;
 514                 sgtmp = sg_next(sgtmp);
 515                 used--;
 516         }
 517         used = nelems - used;
 518
 519         /* Step 4: Create the mappings. */
 520         prot = HV_PCI_MAP_ATTR_READ;
 521         if (direction != DMA_TO_DEVICE)
 522                 prot |= HV_PCI_MAP_ATTR_WRITE;
 523
 524         err = fill_sg(entry, dev, sglist, used, nelems, prot);
 525         if (unlikely(err < 0L))
 526                 goto iommu_map_failed;
 527
 528         return used;
 529
 530 bad:
 531         if (printk_ratelimit())
 532                 WARN_ON(1);
 533         return 0;
 534
 535 iommu_map_failed:
 536         spin_lock_irqsave(&iommu->lock, flags);
 537         arena_free(&iommu->arena, entry, npages);
 538         spin_unlock_irqrestore(&iommu->lock, flags);
 539
 540         return 0;
 541 }
 542
 543 static void dma_4v_unmap_sg(struct device *dev, struct scatterlist *sglist,
 544                             int nelems, enum dma_data_direction direction)
 545 {
 546         struct pci_pbm_info *pbm;
 547         struct iommu *iommu;
 548         unsigned long flags, i, npages;
 549         struct scatterlist *sg, *sgprv;
 550         long entry;
 551         u32 devhandle, bus_addr;
 552
 553         if (unlikely(direction == DMA_NONE)) {
 554                 if (printk_ratelimit())
 555                         WARN_ON(1);
 556         }
 557
 558         iommu = dev->archdata.iommu;
 559         pbm = dev->archdata.host_controller;
 560         devhandle = pbm->devhandle;
 561
 562         bus_addr = sglist->dma_address & IO_PAGE_MASK;
 563         sgprv = NULL;
 564         for_each_sg(sglist, sg, nelems, i) {
 565                 if (sg->dma_length == 0)
 566                         break;
 567
 568                 sgprv = sg;
 569         }
 570
 571         npages = (IO_PAGE_ALIGN(sgprv->dma_address + sgprv->dma_length) -
 572                   bus_addr) >> IO_PAGE_SHIFT;
 573
 574         entry = ((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
 575
 576         spin_lock_irqsave(&iommu->lock, flags);
 577
 578         arena_free(&iommu->arena, entry, npages);
 579
 580         do {
 581                 unsigned long num;
 582
 583                 num = pci_sun4v_iommu_demap(devhandle, HV_PCI_TSBID(0, entry),
 584                                             npages);
 585                 entry += num;
 586                 npages -= num;
 587         } while (npages != 0);
 588
 589         spin_unlock_irqrestore(&iommu->lock, flags);
 590 }
 591
 592 static void dma_4v_sync_single_for_cpu(struct device *dev,
 593                                        dma_addr_t bus_addr, size_t sz,
 594                                        enum dma_data_direction direction)
 595 {
 596         /* Nothing to do... */
 597 }
 598
 599 static void dma_4v_sync_sg_for_cpu(struct device *dev,
 600                                    struct scatterlist *sglist, int nelems,
 601                                    enum dma_data_direction direction)
 602 {
 603         /* Nothing to do... */
 604 }
 605
 606 const struct dma_ops sun4v_dma_ops = {
 607         .alloc_coherent                 = dma_4v_alloc_coherent,
 608         .free_coherent                  = dma_4v_free_coherent,
 609         .map_single                     = dma_4v_map_single,
 610         .unmap_single                   = dma_4v_unmap_single,
 611         .map_sg                         = dma_4v_map_sg,
 612         .unmap_sg                       = dma_4v_unmap_sg,
 613         .sync_single_for_cpu            = dma_4v_sync_single_for_cpu,
 614         .sync_sg_for_cpu                = dma_4v_sync_sg_for_cpu,
 615 };
 616
 617 static void pci_sun4v_scan_bus(struct pci_pbm_info *pbm)
 618 {
 619         struct property *prop;
 620         struct device_node *dp;
 621
 622         dp = pbm->prom_node;
 623         prop = of_find_property(dp, "66mhz-capable", NULL);
 624         pbm->is_66mhz_capable = (prop != NULL);
 625         pbm->pci_bus = pci_scan_one_pbm(pbm);
 626
 627         /* XXX register error interrupt handlers XXX */
 628 }
 629
 630 static unsigned long probe_existing_entries(struct pci_pbm_info *pbm,
 631                                             struct iommu *iommu)
 632 {
 633         struct iommu_arena *arena = &iommu->arena;
 634         unsigned long i, cnt = 0;
 635         u32 devhandle;
 636
 637         devhandle = pbm->devhandle;
 638         for (i = 0; i < arena->limit; i++) {
 639                 unsigned long ret, io_attrs, ra;
 640
 641                 ret = pci_sun4v_iommu_getmap(devhandle,
 642                                              HV_PCI_TSBID(0, i),
 643                                              &io_attrs, &ra);
 644                 if (ret == HV_EOK) {
 645                         if (page_in_phys_avail(ra)) {
 646                                 pci_sun4v_iommu_demap(devhandle,
 647                                                       HV_PCI_TSBID(0, i), 1);
 648                         } else {
 649                                 cnt++;
 650                                 __set_bit(i, arena->map);
 651                         }
 652                 }
 653         }
 654
 655         return cnt;
 656 }
 657
 658 static void pci_sun4v_iommu_init(struct pci_pbm_info *pbm)
 659 {
 660         struct iommu *iommu = pbm->iommu;
 661         struct property *prop;
 662         unsigned long num_tsb_entries, sz, tsbsize;
 663         u32 vdma[2], dma_mask, dma_offset;
 664
 665         prop = of_find_property(pbm->prom_node, "virtual-dma", NULL);
 666         if (prop) {
 667                 u32 *val = prop->value;
 668
 669                 vdma[0] = val[0];
 670                 vdma[1] = val[1];
 671         } else {
 672                 /* No property, use default values. */
 673                 vdma[0] = 0x80000000;
 674                 vdma[1] = 0x80000000;
 675         }
 676
 677         if ((vdma[0] | vdma[1]) & ~IO_PAGE_MASK) {
 678                 prom_printf("PCI-SUN4V: strange virtual-dma[%08x:%08x].\n",
 679                             vdma[0], vdma[1]);
 680                 prom_halt();
 681         };
 682
 683         dma_mask = (roundup_pow_of_two(vdma[1]) - 1UL);
 684         num_tsb_entries = vdma[1] / IO_PAGE_SIZE;
 685         tsbsize = num_tsb_entries * sizeof(iopte_t);
 686
 687         dma_offset = vdma[0];
 688
 689         /* Setup initial software IOMMU state. */
 690         spin_lock_init(&iommu->lock);
 691         iommu->ctx_lowest_free = 1;
 692         iommu->page_table_map_base = dma_offset;
 693         iommu->dma_addr_mask = dma_mask;
 694
 695         /* Allocate and initialize the free area map.  */
 696         sz = (num_tsb_entries + 7) / 8;
 697         sz = (sz + 7UL) & ~7UL;
 698         iommu->arena.map = kzalloc(sz, GFP_KERNEL);
 699         if (!iommu->arena.map) {
 700                 prom_printf("PCI_IOMMU: Error, kmalloc(arena.map) failed.\n");
 701                 prom_halt();
 702         }
 703         iommu->arena.limit = num_tsb_entries;
 704
 705         sz = probe_existing_entries(pbm, iommu);
 706         if (sz)
 707                 printk("%s: Imported %lu TSB entries from OBP\n",
 708                        pbm->name, sz);
 709 }
 710
 711 #ifdef CONFIG_PCI_MSI
 712 struct pci_sun4v_msiq_entry {
 713         u64             version_type;
 714 #define MSIQ_VERSION_MASK               0xffffffff00000000UL
 715 #define MSIQ_VERSION_SHIFT              32
 716 #define MSIQ_TYPE_MASK                  0x00000000000000ffUL
 717 #define MSIQ_TYPE_SHIFT                 0
 718 #define MSIQ_TYPE_NONE                  0x00
 719 #define MSIQ_TYPE_MSG                   0x01
 720 #define MSIQ_TYPE_MSI32                 0x02
 721 #define MSIQ_TYPE_MSI64                 0x03
 722 #define MSIQ_TYPE_INTX                  0x08
 723 #define MSIQ_TYPE_NONE2                 0xff
 724
 725         u64             intx_sysino;
 726         u64             reserved1;
 727         u64             stick;
 728         u64             req_id;  /* bus/device/func */
 729 #define MSIQ_REQID_BUS_MASK             0xff00UL
 730 #define MSIQ_REQID_BUS_SHIFT            8
 731 #define MSIQ_REQID_DEVICE_MASK          0x00f8UL
 732 #define MSIQ_REQID_DEVICE_SHIFT         3
 733 #define MSIQ_REQID_FUNC_MASK            0x0007UL
 734 #define MSIQ_REQID_FUNC_SHIFT           0
 735
 736         u64             msi_address;
 737
 738         /* The format of this value is message type dependent.
 739          * For MSI bits 15:0 are the data from the MSI packet.
 740          * For MSI-X bits 31:0 are the data from the MSI packet.
 741          * For MSG, the message code and message routing code where:
 742          *      bits 39:32 is the bus/device/fn of the msg target-id
 743          *      bits 18:16 is the message routing code
 744          *      bits 7:0 is the message code
 745          * For INTx the low order 2-bits are:
 746          *      00 - INTA
 747          *      01 - INTB
 748          *      10 - INTC
 749          *      11 - INTD
 750          */
 751         u64             msi_data;
 752
 753         u64             reserved2;
 754 };
 755
 756 static int pci_sun4v_get_head(struct pci_pbm_info *pbm, unsigned long msiqid,
 757                               unsigned long *head)
 758 {
 759         unsigned long err, limit;
 760
 761         err = pci_sun4v_msiq_gethead(pbm->devhandle, msiqid, head);
 762         if (unlikely(err))
 763                 return -ENXIO;
 764
 765         limit = pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry);
 766         if (unlikely(*head >= limit))
 767                 return -EFBIG;
 768
 769         return 0;
 770 }
 771
 772 static int pci_sun4v_dequeue_msi(struct pci_pbm_info *pbm,
 773                                  unsigned long msiqid, unsigned long *head,
 774                                  unsigned long *msi)
 775 {
 776         struct pci_sun4v_msiq_entry *ep;
 777         unsigned long err, type;
 778
 779         /* Note: void pointer arithmetic, 'head' is a byte offset  */
 780         ep = (pbm->msi_queues + ((msiqid - pbm->msiq_first) *
 781                                  (pbm->msiq_ent_count *
 782                                   sizeof(struct pci_sun4v_msiq_entry))) +
 783               *head);
 784
 785         if ((ep->version_type & MSIQ_TYPE_MASK) == 0)
 786                 return 0;
 787
 788         type = (ep->version_type & MSIQ_TYPE_MASK) >> MSIQ_TYPE_SHIFT;
 789         if (unlikely(type != MSIQ_TYPE_MSI32 &&
 790                      type != MSIQ_TYPE_MSI64))
 791                 return -EINVAL;
 792
 793         *msi = ep->msi_data;
 794
 795         err = pci_sun4v_msi_setstate(pbm->devhandle,
 796                                      ep->msi_data /* msi_num */,
 797                                      HV_MSISTATE_IDLE);
 798         if (unlikely(err))
 799                 return -ENXIO;
 800
 801         /* Clear the entry.  */
 802         ep->version_type &= ~MSIQ_TYPE_MASK;
 803
 804         (*head) += sizeof(struct pci_sun4v_msiq_entry);
 805         if (*head >=
 806             (pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry)))
 807                 *head = 0;
 808
 809         return 1;
 810 }
 811
 812 static int pci_sun4v_set_head(struct pci_pbm_info *pbm, unsigned long msiqid,
 813                               unsigned long head)
 814 {
 815         unsigned long err;
 816
 817         err = pci_sun4v_msiq_sethead(pbm->devhandle, msiqid, head);
 818         if (unlikely(err))
 819                 return -EINVAL;
 820
 821         return 0;
 822 }
 823
 824 static int pci_sun4v_msi_setup(struct pci_pbm_info *pbm, unsigned long msiqid,
 825                                unsigned long msi, int is_msi64)
 826 {
 827         if (pci_sun4v_msi_setmsiq(pbm->devhandle, msi, msiqid,
 828                                   (is_msi64 ?
 829                                    HV_MSITYPE_MSI64 : HV_MSITYPE_MSI32)))
 830                 return -ENXIO;
 831         if (pci_sun4v_msi_setstate(pbm->devhandle, msi, HV_MSISTATE_IDLE))
 832                 return -ENXIO;
 833         if (pci_sun4v_msi_setvalid(pbm->devhandle, msi, HV_MSIVALID_VALID))
 834                 return -ENXIO;
 835         return 0;
 836 }
 837
 838 static int pci_sun4v_msi_teardown(struct pci_pbm_info *pbm, unsigned long msi)
 839 {
 840         unsigned long err, msiqid;
 841
 842         err = pci_sun4v_msi_getmsiq(pbm->devhandle, msi, &msiqid);
 843         if (err)
 844                 return -ENXIO;
 845
 846         pci_sun4v_msi_setvalid(pbm->devhandle, msi, HV_MSIVALID_INVALID);
 847
 848         return 0;
 849 }
 850
 851 static int pci_sun4v_msiq_alloc(struct pci_pbm_info *pbm)
 852 {
 853         unsigned long q_size, alloc_size, pages, order;
 854         int i;
 855
 856         q_size = pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry);
 857         alloc_size = (pbm->msiq_num * q_size);
 858         order = get_order(alloc_size);
 859         pages = __get_free_pages(GFP_KERNEL | __GFP_COMP, order);
 860         if (pages == 0UL) {
 861                 printk(KERN_ERR "MSI: Cannot allocate MSI queues (o=%lu).\n",
 862                        order);
 863                 return -ENOMEM;
 864         }
 865         memset((char *)pages, 0, PAGE_SIZE << order);
 866         pbm->msi_queues = (void *) pages;
 867
 868         for (i = 0; i < pbm->msiq_num; i++) {
 869                 unsigned long err, base = __pa(pages + (i * q_size));
 870                 unsigned long ret1, ret2;
 871
 872                 err = pci_sun4v_msiq_conf(pbm->devhandle,
 873                                           pbm->msiq_first + i,
 874                                           base, pbm->msiq_ent_count);
 875                 if (err) {
 876                         printk(KERN_ERR "MSI: msiq register fails (err=%lu)\n",
 877                                err);
 878                         goto h_error;
 879                 }
 880
 881                 err = pci_sun4v_msiq_info(pbm->devhandle,
 882                                           pbm->msiq_first + i,
 883                                           &ret1, &ret2);
 884                 if (err) {
 885                         printk(KERN_ERR "MSI: Cannot read msiq (err=%lu)\n",
 886                                err);
 887                         goto h_error;
 888                 }
 889                 if (ret1 != base || ret2 != pbm->msiq_ent_count) {
 890                         printk(KERN_ERR "MSI: Bogus qconf "
 891                                "expected[%lx:%x] got[%lx:%lx]\n",
 892                                base, pbm->msiq_ent_count,
 893                                ret1, ret2);
 894                         goto h_error;
 895                 }
 896         }
 897
 898         return 0;
 899
 900 h_error:
 901         free_pages(pages, order);
 902         return -EINVAL;
 903 }
 904
 905 static void pci_sun4v_msiq_free(struct pci_pbm_info *pbm)
 906 {
 907         unsigned long q_size, alloc_size, pages, order;
 908         int i;
 909
 910         for (i = 0; i < pbm->msiq_num; i++) {
 911                 unsigned long msiqid = pbm->msiq_first + i;
 912
 913                 (void) pci_sun4v_msiq_conf(pbm->devhandle, msiqid, 0UL, 0);
 914         }
 915
 916         q_size = pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry);
 917         alloc_size = (pbm->msiq_num * q_size);
 918         order = get_order(alloc_size);
 919
 920         pages = (unsigned long) pbm->msi_queues;
 921
 922         free_pages(pages, order);
 923
 924         pbm->msi_queues = NULL;
 925 }
 926
 927 static int pci_sun4v_msiq_build_irq(struct pci_pbm_info *pbm,
 928                                     unsigned long msiqid,
 929                                     unsigned long devino)
 930 {
 931         unsigned int virt_irq = sun4v_build_irq(pbm->devhandle, devino);
 932
 933         if (!virt_irq)
 934                 return -ENOMEM;
 935
 936         if (pci_sun4v_msiq_setstate(pbm->devhandle, msiqid, HV_MSIQSTATE_IDLE))
 937                 return -EINVAL;
 938         if (pci_sun4v_msiq_setvalid(pbm->devhandle, msiqid, HV_MSIQ_VALID))
 939                 return -EINVAL;
 940
 941         return virt_irq;
 942 }
 943
 944 static const struct sparc64_msiq_ops pci_sun4v_msiq_ops = {
 945         .get_head       =       pci_sun4v_get_head,
 946         .dequeue_msi    =       pci_sun4v_dequeue_msi,
 947         .set_head       =       pci_sun4v_set_head,
 948         .msi_setup      =       pci_sun4v_msi_setup,
 949         .msi_teardown   =       pci_sun4v_msi_teardown,
 950         .msiq_alloc     =       pci_sun4v_msiq_alloc,
 951         .msiq_free      =       pci_sun4v_msiq_free,
 952         .msiq_build_irq =       pci_sun4v_msiq_build_irq,
 953 };
 954
 955 static void pci_sun4v_msi_init(struct pci_pbm_info *pbm)
 956 {
 957         sparc64_pbm_msi_init(pbm, &pci_sun4v_msiq_ops);
 958 }
 959 #else /* CONFIG_PCI_MSI */
 960 static void pci_sun4v_msi_init(struct pci_pbm_info *pbm)
 961 {
 962 }
 963 #endif /* !(CONFIG_PCI_MSI) */
 964
 965 static void __init pci_sun4v_pbm_init(struct pci_controller_info *p, struct device_node *dp, u32 devhandle)
 966 {
 967         struct pci_pbm_info *pbm;
 968
 969         if (devhandle & 0x40)
 970                 pbm = &p->pbm_B;
 971         else
 972                 pbm = &p->pbm_A;
 973
 974         pbm->next = pci_pbm_root;
 975         pci_pbm_root = pbm;
 976
 977         pbm->scan_bus = pci_sun4v_scan_bus;
 978         pbm->pci_ops = &sun4v_pci_ops;
 979         pbm->config_space_reg_bits = 12;
 980
 981         pbm->index = pci_num_pbms++;
 982
 983         pbm->parent = p;
 984         pbm->prom_node = dp;
 985
 986         pbm->devhandle = devhandle;
 987
 988         pbm->name = dp->full_name;
 989
 990         printk("%s: SUN4V PCI Bus Module\n", pbm->name);
 991
 992         pci_determine_mem_io_space(pbm);
 993
 994         pci_get_pbm_props(pbm);
 995         pci_sun4v_iommu_init(pbm);
 996         pci_sun4v_msi_init(pbm);
 997 }
 998
 999 void __init sun4v_pci_init(struct device_node *dp, char *model_name)
1000 {
1001         static int hvapi_negotiated = 0;
1002         struct pci_controller_info *p;
1003         struct pci_pbm_info *pbm;
1004         struct iommu *iommu;
1005         struct property *prop;
1006         struct linux_prom64_registers *regs;
1007         u32 devhandle;
1008         int i;
1009
1010         if (!hvapi_negotiated++) {
1011                 int err = sun4v_hvapi_register(HV_GRP_PCI,
1012                                                vpci_major,
1013                                                &vpci_minor);
1014
1015                 if (err) {
1016                         prom_printf("SUN4V_PCI: Could not register hvapi, "
1017                                     "err=%d\n", err);
1018                         prom_halt();
1019                 }
1020                 printk("SUN4V_PCI: Registered hvapi major[%lu] minor[%lu]\n",
1021                        vpci_major, vpci_minor);
1022
1023                 dma_ops = &sun4v_dma_ops;
1024         }
1025
1026         prop = of_find_property(dp, "reg", NULL);
1027         regs = prop->value;
1028
1029         devhandle = (regs->phys_addr >> 32UL) & 0x0fffffff;
1030
1031         for (pbm = pci_pbm_root; pbm; pbm = pbm->next) {
1032                 if (pbm->devhandle == (devhandle ^ 0x40)) {
1033                         pci_sun4v_pbm_init(pbm->parent, dp, devhandle);
1034                         return;
1035                 }
1036         }
1037
1038         for_each_possible_cpu(i) {
1039                 unsigned long page = get_zeroed_page(GFP_ATOMIC);
1040
1041                 if (!page)
1042                         goto fatal_memory_error;
1043
1044                 per_cpu(iommu_batch, i).pglist = (u64 *) page;
1045         }
1046
1047         p = kzalloc(sizeof(struct pci_controller_info), GFP_ATOMIC);
1048         if (!p)
1049                 goto fatal_memory_error;
1050
1051         iommu = kzalloc(sizeof(struct iommu), GFP_ATOMIC);
1052         if (!iommu)
1053                 goto fatal_memory_error;
1054
1055         p->pbm_A.iommu = iommu;
1056
1057         iommu = kzalloc(sizeof(struct iommu), GFP_ATOMIC);
1058         if (!iommu)
1059                 goto fatal_memory_error;
1060
1061         p->pbm_B.iommu = iommu;
1062
1063         pci_sun4v_pbm_init(p, dp, devhandle);
1064         return;
1065
1066 fatal_memory_error:
1067         prom_printf("SUN4V_PCI: Fatal memory allocation error.\n");
1068         prom_halt();
1069 }