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) (__pa(sg_virt((SG))))
 369
 370 static long fill_sg(long entry, struct device *dev,
 371                     struct scatterlist *sg,
 372                     int nused, int nelems, unsigned long prot)
 373 {
 374         struct scatterlist *dma_sg = sg;
 375         unsigned long flags;
 376         int i;
 377
 378         local_irq_save(flags);
 379
 380         iommu_batch_start(dev, prot, entry);
 381
 382         for (i = 0; i < nused; i++) {
 383                 unsigned long pteval = ~0UL;
 384                 u32 dma_npages;
 385
 386                 dma_npages = ((dma_sg->dma_address & (IO_PAGE_SIZE - 1UL)) +
 387                               dma_sg->dma_length +
 388                               ((IO_PAGE_SIZE - 1UL))) >> IO_PAGE_SHIFT;
 389                 do {
 390                         unsigned long offset;
 391                         signed int len;
 392
 393                         /* If we are here, we know we have at least one
 394                          * more page to map.  So walk forward until we
 395                          * hit a page crossing, and begin creating new
 396                          * mappings from that spot.
 397                          */
 398                         for (;;) {
 399                                 unsigned long tmp;
 400
 401                                 tmp = SG_ENT_PHYS_ADDRESS(sg);
 402                                 len = sg->length;
 403                                 if (((tmp ^ pteval) >> IO_PAGE_SHIFT) != 0UL) {
 404                                         pteval = tmp & IO_PAGE_MASK;
 405                                         offset = tmp & (IO_PAGE_SIZE - 1UL);
 406                                         break;
 407                                 }
 408                                 if (((tmp ^ (tmp + len - 1UL)) >> IO_PAGE_SHIFT) != 0UL) {
 409                                         pteval = (tmp + IO_PAGE_SIZE) & IO_PAGE_MASK;
 410                                         offset = 0UL;
 411                                         len -= (IO_PAGE_SIZE - (tmp & (IO_PAGE_SIZE - 1UL)));
 412                                         break;
 413                                 }
 414                                 sg = sg_next(sg);
 415                                 nelems--;
 416                         }
 417
 418                         pteval = (pteval & IOPTE_PAGE);
 419                         while (len > 0) {
 420                                 long err;
 421
 422                                 err = iommu_batch_add(pteval);
 423                                 if (unlikely(err < 0L))
 424                                         goto iommu_map_failed;
 425
 426                                 pteval += IO_PAGE_SIZE;
 427                                 len -= (IO_PAGE_SIZE - offset);
 428                                 offset = 0;
 429                                 dma_npages--;
 430                         }
 431
 432                         pteval = (pteval & IOPTE_PAGE) + len;
 433                         sg = sg_next(sg);
 434                         nelems--;
 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 (nelems &&
 441                                (pteval << (64 - IO_PAGE_SHIFT)) != 0UL &&
 442                                (pteval == SG_ENT_PHYS_ADDRESS(sg)) &&
 443                                ((pteval ^
 444                                  (SG_ENT_PHYS_ADDRESS(sg) + sg->length - 1UL)) >> IO_PAGE_SHIFT) == 0UL) {
 445                                 pteval += sg->length;
 446                                 sg = sg_next(sg);
 447                                 nelems--;
 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, sg_virt(sglist),
 480                                           sglist->length, direction);
 481                 if (unlikely(sglist->dma_address == DMA_ERROR_CODE))
 482                         return 0;
 483                 sglist->dma_length = sglist->length;
 484                 return 1;
 485         }
 486
 487         iommu = dev->archdata.iommu;
 488
 489         if (unlikely(direction == DMA_NONE))
 490                 goto bad;
 491
 492         /* Step 1: Prepare scatter list. */
 493         npages = prepare_sg(sglist, nelems);
 494
 495         /* Step 2: Allocate a cluster and context, if necessary. */
 496         spin_lock_irqsave(&iommu->lock, flags);
 497         entry = arena_alloc(&iommu->arena, npages);
 498         spin_unlock_irqrestore(&iommu->lock, flags);
 499
 500         if (unlikely(entry < 0L))
 501                 goto bad;
 502
 503         dma_base = iommu->page_table_map_base +
 504                 (entry << IO_PAGE_SHIFT);
 505
 506         /* Step 3: Normalize DMA addresses. */
 507         used = nelems;
 508
 509         sgtmp = sglist;
 510         while (used && sgtmp->dma_length) {
 511                 sgtmp->dma_address += dma_base;
 512                 sgtmp = sg_next(sgtmp);
 513                 used--;
 514         }
 515         used = nelems - used;
 516
 517         /* Step 4: Create the mappings. */
 518         prot = HV_PCI_MAP_ATTR_READ;
 519         if (direction != DMA_TO_DEVICE)
 520                 prot |= HV_PCI_MAP_ATTR_WRITE;
 521
 522         err = fill_sg(entry, dev, sglist, used, nelems, prot);
 523         if (unlikely(err < 0L))
 524                 goto iommu_map_failed;
 525
 526         return used;
 527
 528 bad:
 529         if (printk_ratelimit())
 530                 WARN_ON(1);
 531         return 0;
 532
 533 iommu_map_failed:
 534         spin_lock_irqsave(&iommu->lock, flags);
 535         arena_free(&iommu->arena, entry, npages);
 536         spin_unlock_irqrestore(&iommu->lock, flags);
 537
 538         return 0;
 539 }
 540
 541 static void dma_4v_unmap_sg(struct device *dev, struct scatterlist *sglist,
 542                             int nelems, enum dma_data_direction direction)
 543 {
 544         struct pci_pbm_info *pbm;
 545         struct iommu *iommu;
 546         unsigned long flags, i, npages;
 547         struct scatterlist *sg, *sgprv;
 548         long entry;
 549         u32 devhandle, bus_addr;
 550
 551         if (unlikely(direction == DMA_NONE)) {
 552                 if (printk_ratelimit())
 553                         WARN_ON(1);
 554         }
 555
 556         iommu = dev->archdata.iommu;
 557         pbm = dev->archdata.host_controller;
 558         devhandle = pbm->devhandle;
 559
 560         bus_addr = sglist->dma_address & IO_PAGE_MASK;
 561         sgprv = NULL;
 562         for_each_sg(sglist, sg, nelems, i) {
 563                 if (sg->dma_length == 0)
 564                         break;
 565
 566                 sgprv = sg;
 567         }
 568
 569         npages = (IO_PAGE_ALIGN(sgprv->dma_address + sgprv->dma_length) -
 570                   bus_addr) >> IO_PAGE_SHIFT;
 571
 572         entry = ((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
 573
 574         spin_lock_irqsave(&iommu->lock, flags);
 575
 576         arena_free(&iommu->arena, entry, npages);
 577
 578         do {
 579                 unsigned long num;
 580
 581                 num = pci_sun4v_iommu_demap(devhandle, HV_PCI_TSBID(0, entry),
 582                                             npages);
 583                 entry += num;
 584                 npages -= num;
 585         } while (npages != 0);
 586
 587         spin_unlock_irqrestore(&iommu->lock, flags);
 588 }
 589
 590 static void dma_4v_sync_single_for_cpu(struct device *dev,
 591                                        dma_addr_t bus_addr, size_t sz,
 592                                        enum dma_data_direction direction)
 593 {
 594         /* Nothing to do... */
 595 }
 596
 597 static void dma_4v_sync_sg_for_cpu(struct device *dev,
 598                                    struct scatterlist *sglist, int nelems,
 599                                    enum dma_data_direction direction)
 600 {
 601         /* Nothing to do... */
 602 }
 603
 604 const struct dma_ops sun4v_dma_ops = {
 605         .alloc_coherent                 = dma_4v_alloc_coherent,
 606         .free_coherent                  = dma_4v_free_coherent,
 607         .map_single                     = dma_4v_map_single,
 608         .unmap_single                   = dma_4v_unmap_single,
 609         .map_sg                         = dma_4v_map_sg,
 610         .unmap_sg                       = dma_4v_unmap_sg,
 611         .sync_single_for_cpu            = dma_4v_sync_single_for_cpu,
 612         .sync_sg_for_cpu                = dma_4v_sync_sg_for_cpu,
 613 };
 614
 615 static void __init pci_sun4v_scan_bus(struct pci_pbm_info *pbm)
 616 {
 617         struct property *prop;
 618         struct device_node *dp;
 619
 620         dp = pbm->prom_node;
 621         prop = of_find_property(dp, "66mhz-capable", NULL);
 622         pbm->is_66mhz_capable = (prop != NULL);
 623         pbm->pci_bus = pci_scan_one_pbm(pbm);
 624
 625         /* XXX register error interrupt handlers XXX */
 626 }
 627
 628 static unsigned long probe_existing_entries(struct pci_pbm_info *pbm,
 629                                             struct iommu *iommu)
 630 {
 631         struct iommu_arena *arena = &iommu->arena;
 632         unsigned long i, cnt = 0;
 633         u32 devhandle;
 634
 635         devhandle = pbm->devhandle;
 636         for (i = 0; i < arena->limit; i++) {
 637                 unsigned long ret, io_attrs, ra;
 638
 639                 ret = pci_sun4v_iommu_getmap(devhandle,
 640                                              HV_PCI_TSBID(0, i),
 641                                              &io_attrs, &ra);
 642                 if (ret == HV_EOK) {
 643                         if (page_in_phys_avail(ra)) {
 644                                 pci_sun4v_iommu_demap(devhandle,
 645                                                       HV_PCI_TSBID(0, i), 1);
 646                         } else {
 647                                 cnt++;
 648                                 __set_bit(i, arena->map);
 649                         }
 650                 }
 651         }
 652
 653         return cnt;
 654 }
 655
 656 static void pci_sun4v_iommu_init(struct pci_pbm_info *pbm)
 657 {
 658         struct iommu *iommu = pbm->iommu;
 659         struct property *prop;
 660         unsigned long num_tsb_entries, sz, tsbsize;
 661         u32 vdma[2], dma_mask, dma_offset;
 662
 663         prop = of_find_property(pbm->prom_node, "virtual-dma", NULL);
 664         if (prop) {
 665                 u32 *val = prop->value;
 666
 667                 vdma[0] = val[0];
 668                 vdma[1] = val[1];
 669         } else {
 670                 /* No property, use default values. */
 671                 vdma[0] = 0x80000000;
 672                 vdma[1] = 0x80000000;
 673         }
 674
 675         if ((vdma[0] | vdma[1]) & ~IO_PAGE_MASK) {
 676                 prom_printf("PCI-SUN4V: strange virtual-dma[%08x:%08x].\n",
 677                             vdma[0], vdma[1]);
 678                 prom_halt();
 679         };
 680
 681         dma_mask = (roundup_pow_of_two(vdma[1]) - 1UL);
 682         num_tsb_entries = vdma[1] / IO_PAGE_SIZE;
 683         tsbsize = num_tsb_entries * sizeof(iopte_t);
 684
 685         dma_offset = vdma[0];
 686
 687         /* Setup initial software IOMMU state. */
 688         spin_lock_init(&iommu->lock);
 689         iommu->ctx_lowest_free = 1;
 690         iommu->page_table_map_base = dma_offset;
 691         iommu->dma_addr_mask = dma_mask;
 692
 693         /* Allocate and initialize the free area map.  */
 694         sz = (num_tsb_entries + 7) / 8;
 695         sz = (sz + 7UL) & ~7UL;
 696         iommu->arena.map = kzalloc(sz, GFP_KERNEL);
 697         if (!iommu->arena.map) {
 698                 prom_printf("PCI_IOMMU: Error, kmalloc(arena.map) failed.\n");
 699                 prom_halt();
 700         }
 701         iommu->arena.limit = num_tsb_entries;
 702
 703         sz = probe_existing_entries(pbm, iommu);
 704         if (sz)
 705                 printk("%s: Imported %lu TSB entries from OBP\n",
 706                        pbm->name, sz);
 707 }
 708
 709 #ifdef CONFIG_PCI_MSI
 710 struct pci_sun4v_msiq_entry {
 711         u64             version_type;
 712 #define MSIQ_VERSION_MASK               0xffffffff00000000UL
 713 #define MSIQ_VERSION_SHIFT              32
 714 #define MSIQ_TYPE_MASK                  0x00000000000000ffUL
 715 #define MSIQ_TYPE_SHIFT                 0
 716 #define MSIQ_TYPE_NONE                  0x00
 717 #define MSIQ_TYPE_MSG                   0x01
 718 #define MSIQ_TYPE_MSI32                 0x02
 719 #define MSIQ_TYPE_MSI64                 0x03
 720 #define MSIQ_TYPE_INTX                  0x08
 721 #define MSIQ_TYPE_NONE2                 0xff
 722
 723         u64             intx_sysino;
 724         u64             reserved1;
 725         u64             stick;
 726         u64             req_id;  /* bus/device/func */
 727 #define MSIQ_REQID_BUS_MASK             0xff00UL
 728 #define MSIQ_REQID_BUS_SHIFT            8
 729 #define MSIQ_REQID_DEVICE_MASK          0x00f8UL
 730 #define MSIQ_REQID_DEVICE_SHIFT         3
 731 #define MSIQ_REQID_FUNC_MASK            0x0007UL
 732 #define MSIQ_REQID_FUNC_SHIFT           0
 733
 734         u64             msi_address;
 735
 736         /* The format of this value is message type dependent.
 737          * For MSI bits 15:0 are the data from the MSI packet.
 738          * For MSI-X bits 31:0 are the data from the MSI packet.
 739          * For MSG, the message code and message routing code where:
 740          *      bits 39:32 is the bus/device/fn of the msg target-id
 741          *      bits 18:16 is the message routing code
 742          *      bits 7:0 is the message code
 743          * For INTx the low order 2-bits are:
 744          *      00 - INTA
 745          *      01 - INTB
 746          *      10 - INTC
 747          *      11 - INTD
 748          */
 749         u64             msi_data;
 750
 751         u64             reserved2;
 752 };
 753
 754 static int pci_sun4v_get_head(struct pci_pbm_info *pbm, unsigned long msiqid,
 755                               unsigned long *head)
 756 {
 757         unsigned long err, limit;
 758
 759         err = pci_sun4v_msiq_gethead(pbm->devhandle, msiqid, head);
 760         if (unlikely(err))
 761                 return -ENXIO;
 762
 763         limit = pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry);
 764         if (unlikely(*head >= limit))
 765                 return -EFBIG;
 766
 767         return 0;
 768 }
 769
 770 static int pci_sun4v_dequeue_msi(struct pci_pbm_info *pbm,
 771                                  unsigned long msiqid, unsigned long *head,
 772                                  unsigned long *msi)
 773 {
 774         struct pci_sun4v_msiq_entry *ep;
 775         unsigned long err, type;
 776
 777         /* Note: void pointer arithmetic, 'head' is a byte offset  */
 778         ep = (pbm->msi_queues + ((msiqid - pbm->msiq_first) *
 779                                  (pbm->msiq_ent_count *
 780                                   sizeof(struct pci_sun4v_msiq_entry))) +
 781               *head);
 782
 783         if ((ep->version_type & MSIQ_TYPE_MASK) == 0)
 784                 return 0;
 785
 786         type = (ep->version_type & MSIQ_TYPE_MASK) >> MSIQ_TYPE_SHIFT;
 787         if (unlikely(type != MSIQ_TYPE_MSI32 &&
 788                      type != MSIQ_TYPE_MSI64))
 789                 return -EINVAL;
 790
 791         *msi = ep->msi_data;
 792
 793         err = pci_sun4v_msi_setstate(pbm->devhandle,
 794                                      ep->msi_data /* msi_num */,
 795                                      HV_MSISTATE_IDLE);
 796         if (unlikely(err))
 797                 return -ENXIO;
 798
 799         /* Clear the entry.  */
 800         ep->version_type &= ~MSIQ_TYPE_MASK;
 801
 802         (*head) += sizeof(struct pci_sun4v_msiq_entry);
 803         if (*head >=
 804             (pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry)))
 805                 *head = 0;
 806
 807         return 1;
 808 }
 809
 810 static int pci_sun4v_set_head(struct pci_pbm_info *pbm, unsigned long msiqid,
 811                               unsigned long head)
 812 {
 813         unsigned long err;
 814
 815         err = pci_sun4v_msiq_sethead(pbm->devhandle, msiqid, head);
 816         if (unlikely(err))
 817                 return -EINVAL;
 818
 819         return 0;
 820 }
 821
 822 static int pci_sun4v_msi_setup(struct pci_pbm_info *pbm, unsigned long msiqid,
 823                                unsigned long msi, int is_msi64)
 824 {
 825         if (pci_sun4v_msi_setmsiq(pbm->devhandle, msi, msiqid,
 826                                   (is_msi64 ?
 827                                    HV_MSITYPE_MSI64 : HV_MSITYPE_MSI32)))
 828                 return -ENXIO;
 829         if (pci_sun4v_msi_setstate(pbm->devhandle, msi, HV_MSISTATE_IDLE))
 830                 return -ENXIO;
 831         if (pci_sun4v_msi_setvalid(pbm->devhandle, msi, HV_MSIVALID_VALID))
 832                 return -ENXIO;
 833         return 0;
 834 }
 835
 836 static int pci_sun4v_msi_teardown(struct pci_pbm_info *pbm, unsigned long msi)
 837 {
 838         unsigned long err, msiqid;
 839
 840         err = pci_sun4v_msi_getmsiq(pbm->devhandle, msi, &msiqid);
 841         if (err)
 842                 return -ENXIO;
 843
 844         pci_sun4v_msi_setvalid(pbm->devhandle, msi, HV_MSIVALID_INVALID);
 845
 846         return 0;
 847 }
 848
 849 static int pci_sun4v_msiq_alloc(struct pci_pbm_info *pbm)
 850 {
 851         unsigned long q_size, alloc_size, pages, order;
 852         int i;
 853
 854         q_size = pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry);
 855         alloc_size = (pbm->msiq_num * q_size);
 856         order = get_order(alloc_size);
 857         pages = __get_free_pages(GFP_KERNEL | __GFP_COMP, order);
 858         if (pages == 0UL) {
 859                 printk(KERN_ERR "MSI: Cannot allocate MSI queues (o=%lu).\n",
 860                        order);
 861                 return -ENOMEM;
 862         }
 863         memset((char *)pages, 0, PAGE_SIZE << order);
 864         pbm->msi_queues = (void *) pages;
 865
 866         for (i = 0; i < pbm->msiq_num; i++) {
 867                 unsigned long err, base = __pa(pages + (i * q_size));
 868                 unsigned long ret1, ret2;
 869
 870                 err = pci_sun4v_msiq_conf(pbm->devhandle,
 871                                           pbm->msiq_first + i,
 872                                           base, pbm->msiq_ent_count);
 873                 if (err) {
 874                         printk(KERN_ERR "MSI: msiq register fails (err=%lu)\n",
 875                                err);
 876                         goto h_error;
 877                 }
 878
 879                 err = pci_sun4v_msiq_info(pbm->devhandle,
 880                                           pbm->msiq_first + i,
 881                                           &ret1, &ret2);
 882                 if (err) {
 883                         printk(KERN_ERR "MSI: Cannot read msiq (err=%lu)\n",
 884                                err);
 885                         goto h_error;
 886                 }
 887                 if (ret1 != base || ret2 != pbm->msiq_ent_count) {
 888                         printk(KERN_ERR "MSI: Bogus qconf "
 889                                "expected[%lx:%x] got[%lx:%lx]\n",
 890                                base, pbm->msiq_ent_count,
 891                                ret1, ret2);
 892                         goto h_error;
 893                 }
 894         }
 895
 896         return 0;
 897
 898 h_error:
 899         free_pages(pages, order);
 900         return -EINVAL;
 901 }
 902
 903 static void pci_sun4v_msiq_free(struct pci_pbm_info *pbm)
 904 {
 905         unsigned long q_size, alloc_size, pages, order;
 906         int i;
 907
 908         for (i = 0; i < pbm->msiq_num; i++) {
 909                 unsigned long msiqid = pbm->msiq_first + i;
 910
 911                 (void) pci_sun4v_msiq_conf(pbm->devhandle, msiqid, 0UL, 0);
 912         }
 913
 914         q_size = pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry);
 915         alloc_size = (pbm->msiq_num * q_size);
 916         order = get_order(alloc_size);
 917
 918         pages = (unsigned long) pbm->msi_queues;
 919
 920         free_pages(pages, order);
 921
 922         pbm->msi_queues = NULL;
 923 }
 924
 925 static int pci_sun4v_msiq_build_irq(struct pci_pbm_info *pbm,
 926                                     unsigned long msiqid,
 927                                     unsigned long devino)
 928 {
 929         unsigned int virt_irq = sun4v_build_irq(pbm->devhandle, devino);
 930
 931         if (!virt_irq)
 932                 return -ENOMEM;
 933
 934         if (pci_sun4v_msiq_setstate(pbm->devhandle, msiqid, HV_MSIQSTATE_IDLE))
 935                 return -EINVAL;
 936         if (pci_sun4v_msiq_setvalid(pbm->devhandle, msiqid, HV_MSIQ_VALID))
 937                 return -EINVAL;
 938
 939         return virt_irq;
 940 }
 941
 942 static const struct sparc64_msiq_ops pci_sun4v_msiq_ops = {
 943         .get_head       =       pci_sun4v_get_head,
 944         .dequeue_msi    =       pci_sun4v_dequeue_msi,
 945         .set_head       =       pci_sun4v_set_head,
 946         .msi_setup      =       pci_sun4v_msi_setup,
 947         .msi_teardown   =       pci_sun4v_msi_teardown,
 948         .msiq_alloc     =       pci_sun4v_msiq_alloc,
 949         .msiq_free      =       pci_sun4v_msiq_free,
 950         .msiq_build_irq =       pci_sun4v_msiq_build_irq,
 951 };
 952
 953 static void pci_sun4v_msi_init(struct pci_pbm_info *pbm)
 954 {
 955         sparc64_pbm_msi_init(pbm, &pci_sun4v_msiq_ops);
 956 }
 957 #else /* CONFIG_PCI_MSI */
 958 static void pci_sun4v_msi_init(struct pci_pbm_info *pbm)
 959 {
 960 }
 961 #endif /* !(CONFIG_PCI_MSI) */
 962
 963 static void __init pci_sun4v_pbm_init(struct pci_controller_info *p,
 964                                       struct device_node *dp, u32 devhandle)
 965 {
 966         struct pci_pbm_info *pbm;
 967
 968         if (devhandle & 0x40)
 969                 pbm = &p->pbm_B;
 970         else
 971                 pbm = &p->pbm_A;
 972
 973         pbm->next = pci_pbm_root;
 974         pci_pbm_root = pbm;
 975
 976         pbm->scan_bus = pci_sun4v_scan_bus;
 977         pbm->pci_ops = &sun4v_pci_ops;
 978         pbm->config_space_reg_bits = 12;
 979
 980         pbm->index = pci_num_pbms++;
 981
 982         pbm->parent = p;
 983         pbm->prom_node = dp;
 984
 985         pbm->devhandle = devhandle;
 986
 987         pbm->name = dp->full_name;
 988
 989         printk("%s: SUN4V PCI Bus Module\n", pbm->name);
 990
 991         pci_determine_mem_io_space(pbm);
 992
 993         pci_get_pbm_props(pbm);
 994         pci_sun4v_iommu_init(pbm);
 995         pci_sun4v_msi_init(pbm);
 996 }
 997
 998 void __init sun4v_pci_init(struct device_node *dp, char *model_name)
 999 {
1000         static int hvapi_negotiated = 0;
1001         struct pci_controller_info *p;
1002         struct pci_pbm_info *pbm;
1003         struct iommu *iommu;
1004         struct property *prop;
1005         struct linux_prom64_registers *regs;
1006         u32 devhandle;
1007         int i;
1008
1009         if (!hvapi_negotiated++) {
1010                 int err = sun4v_hvapi_register(HV_GRP_PCI,
1011                                                vpci_major,
1012                                                &vpci_minor);
1013
1014                 if (err) {
1015                         prom_printf("SUN4V_PCI: Could not register hvapi, "
1016                                     "err=%d\n", err);
1017                         prom_halt();
1018                 }
1019                 printk("SUN4V_PCI: Registered hvapi major[%lu] minor[%lu]\n",
1020                        vpci_major, vpci_minor);
1021
1022                 dma_ops = &sun4v_dma_ops;
1023         }
1024
1025         prop = of_find_property(dp, "reg", NULL);
1026         if (!prop) {
1027                 prom_printf("SUN4V_PCI: Could not find config registers\n");
1028                 prom_halt();
1029         }
1030         regs = prop->value;
1031
1032         devhandle = (regs->phys_addr >> 32UL) & 0x0fffffff;
1033
1034         for (pbm = pci_pbm_root; pbm; pbm = pbm->next) {
1035                 if (pbm->devhandle == (devhandle ^ 0x40)) {
1036                         pci_sun4v_pbm_init(pbm->parent, dp, devhandle);
1037                         return;
1038                 }
1039         }
1040
1041         for_each_possible_cpu(i) {
1042                 unsigned long page = get_zeroed_page(GFP_ATOMIC);
1043
1044                 if (!page)
1045                         goto fatal_memory_error;
1046
1047                 per_cpu(iommu_batch, i).pglist = (u64 *) page;
1048         }
1049
1050         p = kzalloc(sizeof(struct pci_controller_info), GFP_ATOMIC);
1051         if (!p)
1052                 goto fatal_memory_error;
1053
1054         iommu = kzalloc(sizeof(struct iommu), GFP_ATOMIC);
1055         if (!iommu)
1056                 goto fatal_memory_error;
1057
1058         p->pbm_A.iommu = iommu;
1059
1060         iommu = kzalloc(sizeof(struct iommu), GFP_ATOMIC);
1061         if (!iommu)
1062                 goto fatal_memory_error;
1063
1064         p->pbm_B.iommu = iommu;
1065
1066         pci_sun4v_pbm_init(p, dp, devhandle);
1067         return;
1068
1069 fatal_memory_error:
1070         prom_printf("SUN4V_PCI: Fatal memory allocation error.\n");
1071         prom_halt();
1072 }