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 + 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++;
 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++;
 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 (sg < sg_end &&
 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++;
 447                         }
 448                         if ((pteval << (64 - IO_PAGE_SHIFT)) == 0UL)
 449                                 pteval = ~0UL;
 450                 } while (dma_npages != 0);
 451                 dma_sg++;
 452         }
 453
 454         if (unlikely(iommu_batch_end() < 0L))
 455                 goto iommu_map_failed;
 456
 457         local_irq_restore(flags);
 458         return 0;
 459
 460 iommu_map_failed:
 461         local_irq_restore(flags);
 462         return -1L;
 463 }
 464
 465 static int dma_4v_map_sg(struct device *dev, struct scatterlist *sglist,
 466                          int nelems, enum dma_data_direction direction)
 467 {
 468         struct iommu *iommu;
 469         unsigned long flags, npages, prot;
 470         u32 dma_base;
 471         struct scatterlist *sgtmp;
 472         long entry, err;
 473         int used;
 474
 475         /* Fast path single entry scatterlists. */
 476         if (nelems == 1) {
 477                 sglist->dma_address =
 478                         dma_4v_map_single(dev,
 479                                           (page_address(sglist->page) +
 480                                            sglist->offset),
 481                                           sglist->length, direction);
 482                 if (unlikely(sglist->dma_address == DMA_ERROR_CODE))
 483                         return 0;
 484                 sglist->dma_length = sglist->length;
 485                 return 1;
 486         }
 487
 488         iommu = dev->archdata.iommu;
 489
 490         if (unlikely(direction == DMA_NONE))
 491                 goto bad;
 492
 493         /* Step 1: Prepare scatter list. */
 494         npages = prepare_sg(sglist, nelems);
 495
 496         /* Step 2: Allocate a cluster and context, if necessary. */
 497         spin_lock_irqsave(&iommu->lock, flags);
 498         entry = arena_alloc(&iommu->arena, npages);
 499         spin_unlock_irqrestore(&iommu->lock, flags);
 500
 501         if (unlikely(entry < 0L))
 502                 goto bad;
 503
 504         dma_base = iommu->page_table_map_base +
 505                 (entry << IO_PAGE_SHIFT);
 506
 507         /* Step 3: Normalize DMA addresses. */
 508         used = nelems;
 509
 510         sgtmp = sglist;
 511         while (used && sgtmp->dma_length) {
 512                 sgtmp->dma_address += dma_base;
 513                 sgtmp++;
 514                 used--;
 515         }
 516         used = nelems - used;
 517
 518         /* Step 4: Create the mappings. */
 519         prot = HV_PCI_MAP_ATTR_READ;
 520         if (direction != DMA_TO_DEVICE)
 521                 prot |= HV_PCI_MAP_ATTR_WRITE;
 522
 523         err = fill_sg(entry, dev, sglist, used, nelems, prot);
 524         if (unlikely(err < 0L))
 525                 goto iommu_map_failed;
 526
 527         return used;
 528
 529 bad:
 530         if (printk_ratelimit())
 531                 WARN_ON(1);
 532         return 0;
 533
 534 iommu_map_failed:
 535         spin_lock_irqsave(&iommu->lock, flags);
 536         arena_free(&iommu->arena, entry, npages);
 537         spin_unlock_irqrestore(&iommu->lock, flags);
 538
 539         return 0;
 540 }
 541
 542 static void dma_4v_unmap_sg(struct device *dev, struct scatterlist *sglist,
 543                             int nelems, enum dma_data_direction direction)
 544 {
 545         struct pci_pbm_info *pbm;
 546         struct iommu *iommu;
 547         unsigned long flags, i, npages;
 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
 562         for (i = 1; i < nelems; i++)
 563                 if (sglist[i].dma_length == 0)
 564                         break;
 565         i--;
 566         npages = (IO_PAGE_ALIGN(sglist[i].dma_address + sglist[i].dma_length) -
 567                   bus_addr) >> IO_PAGE_SHIFT;
 568
 569         entry = ((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
 570
 571         spin_lock_irqsave(&iommu->lock, flags);
 572
 573         arena_free(&iommu->arena, entry, npages);
 574
 575         do {
 576                 unsigned long num;
 577
 578                 num = pci_sun4v_iommu_demap(devhandle, HV_PCI_TSBID(0, entry),
 579                                             npages);
 580                 entry += num;
 581                 npages -= num;
 582         } while (npages != 0);
 583
 584         spin_unlock_irqrestore(&iommu->lock, flags);
 585 }
 586
 587 static void dma_4v_sync_single_for_cpu(struct device *dev,
 588                                        dma_addr_t bus_addr, size_t sz,
 589                                        enum dma_data_direction direction)
 590 {
 591         /* Nothing to do... */
 592 }
 593
 594 static void dma_4v_sync_sg_for_cpu(struct device *dev,
 595                                    struct scatterlist *sglist, int nelems,
 596                                    enum dma_data_direction direction)
 597 {
 598         /* Nothing to do... */
 599 }
 600
 601 const struct dma_ops sun4v_dma_ops = {
 602         .alloc_coherent                 = dma_4v_alloc_coherent,
 603         .free_coherent                  = dma_4v_free_coherent,
 604         .map_single                     = dma_4v_map_single,
 605         .unmap_single                   = dma_4v_unmap_single,
 606         .map_sg                         = dma_4v_map_sg,
 607         .unmap_sg                       = dma_4v_unmap_sg,
 608         .sync_single_for_cpu            = dma_4v_sync_single_for_cpu,
 609         .sync_sg_for_cpu                = dma_4v_sync_sg_for_cpu,
 610 };
 611
 612 static void pci_sun4v_scan_bus(struct pci_pbm_info *pbm)
 613 {
 614         struct property *prop;
 615         struct device_node *dp;
 616
 617         dp = pbm->prom_node;
 618         prop = of_find_property(dp, "66mhz-capable", NULL);
 619         pbm->is_66mhz_capable = (prop != NULL);
 620         pbm->pci_bus = pci_scan_one_pbm(pbm);
 621
 622         /* XXX register error interrupt handlers XXX */
 623 }
 624
 625 static unsigned long probe_existing_entries(struct pci_pbm_info *pbm,
 626                                             struct iommu *iommu)
 627 {
 628         struct iommu_arena *arena = &iommu->arena;
 629         unsigned long i, cnt = 0;
 630         u32 devhandle;
 631
 632         devhandle = pbm->devhandle;
 633         for (i = 0; i < arena->limit; i++) {
 634                 unsigned long ret, io_attrs, ra;
 635
 636                 ret = pci_sun4v_iommu_getmap(devhandle,
 637                                              HV_PCI_TSBID(0, i),
 638                                              &io_attrs, &ra);
 639                 if (ret == HV_EOK) {
 640                         if (page_in_phys_avail(ra)) {
 641                                 pci_sun4v_iommu_demap(devhandle,
 642                                                       HV_PCI_TSBID(0, i), 1);
 643                         } else {
 644                                 cnt++;
 645                                 __set_bit(i, arena->map);
 646                         }
 647                 }
 648         }
 649
 650         return cnt;
 651 }
 652
 653 static void pci_sun4v_iommu_init(struct pci_pbm_info *pbm)
 654 {
 655         struct iommu *iommu = pbm->iommu;
 656         struct property *prop;
 657         unsigned long num_tsb_entries, sz, tsbsize;
 658         u32 vdma[2], dma_mask, dma_offset;
 659
 660         prop = of_find_property(pbm->prom_node, "virtual-dma", NULL);
 661         if (prop) {
 662                 u32 *val = prop->value;
 663
 664                 vdma[0] = val[0];
 665                 vdma[1] = val[1];
 666         } else {
 667                 /* No property, use default values. */
 668                 vdma[0] = 0x80000000;
 669                 vdma[1] = 0x80000000;
 670         }
 671
 672         if ((vdma[0] | vdma[1]) & ~IO_PAGE_MASK) {
 673                 prom_printf("PCI-SUN4V: strange virtual-dma[%08x:%08x].\n",
 674                             vdma[0], vdma[1]);
 675                 prom_halt();
 676         };
 677
 678         dma_mask = (roundup_pow_of_two(vdma[1]) - 1UL);
 679         num_tsb_entries = vdma[1] / IO_PAGE_SIZE;
 680         tsbsize = num_tsb_entries * sizeof(iopte_t);
 681
 682         dma_offset = vdma[0];
 683
 684         /* Setup initial software IOMMU state. */
 685         spin_lock_init(&iommu->lock);
 686         iommu->ctx_lowest_free = 1;
 687         iommu->page_table_map_base = dma_offset;
 688         iommu->dma_addr_mask = dma_mask;
 689
 690         /* Allocate and initialize the free area map.  */
 691         sz = (num_tsb_entries + 7) / 8;
 692         sz = (sz + 7UL) & ~7UL;
 693         iommu->arena.map = kzalloc(sz, GFP_KERNEL);
 694         if (!iommu->arena.map) {
 695                 prom_printf("PCI_IOMMU: Error, kmalloc(arena.map) failed.\n");
 696                 prom_halt();
 697         }
 698         iommu->arena.limit = num_tsb_entries;
 699
 700         sz = probe_existing_entries(pbm, iommu);
 701         if (sz)
 702                 printk("%s: Imported %lu TSB entries from OBP\n",
 703                        pbm->name, sz);
 704 }
 705
 706 #ifdef CONFIG_PCI_MSI
 707 struct pci_sun4v_msiq_entry {
 708         u64             version_type;
 709 #define MSIQ_VERSION_MASK               0xffffffff00000000UL
 710 #define MSIQ_VERSION_SHIFT              32
 711 #define MSIQ_TYPE_MASK                  0x00000000000000ffUL
 712 #define MSIQ_TYPE_SHIFT                 0
 713 #define MSIQ_TYPE_NONE                  0x00
 714 #define MSIQ_TYPE_MSG                   0x01
 715 #define MSIQ_TYPE_MSI32                 0x02
 716 #define MSIQ_TYPE_MSI64                 0x03
 717 #define MSIQ_TYPE_INTX                  0x08
 718 #define MSIQ_TYPE_NONE2                 0xff
 719
 720         u64             intx_sysino;
 721         u64             reserved1;
 722         u64             stick;
 723         u64             req_id;  /* bus/device/func */
 724 #define MSIQ_REQID_BUS_MASK             0xff00UL
 725 #define MSIQ_REQID_BUS_SHIFT            8
 726 #define MSIQ_REQID_DEVICE_MASK          0x00f8UL
 727 #define MSIQ_REQID_DEVICE_SHIFT         3
 728 #define MSIQ_REQID_FUNC_MASK            0x0007UL
 729 #define MSIQ_REQID_FUNC_SHIFT           0
 730
 731         u64             msi_address;
 732
 733         /* The format of this value is message type dependent.
 734          * For MSI bits 15:0 are the data from the MSI packet.
 735          * For MSI-X bits 31:0 are the data from the MSI packet.
 736          * For MSG, the message code and message routing code where:
 737          *      bits 39:32 is the bus/device/fn of the msg target-id
 738          *      bits 18:16 is the message routing code
 739          *      bits 7:0 is the message code
 740          * For INTx the low order 2-bits are:
 741          *      00 - INTA
 742          *      01 - INTB
 743          *      10 - INTC
 744          *      11 - INTD
 745          */
 746         u64             msi_data;
 747
 748         u64             reserved2;
 749 };
 750
 751 /* For now this just runs as a pre-handler for the real interrupt handler.
 752  * So we just walk through the queue and ACK all the entries, update the
 753  * head pointer, and return.
 754  *
 755  * In the longer term it would be nice to do something more integrated
 756  * wherein we can pass in some of this MSI info to the drivers.  This
 757  * would be most useful for PCIe fabric error messages, although we could
 758  * invoke those directly from the loop here in order to pass the info around.
 759  */
 760 static void pci_sun4v_msi_prehandler(unsigned int ino, void *data1, void *data2)
 761 {
 762         struct pci_pbm_info *pbm = data1;
 763         struct pci_sun4v_msiq_entry *base, *ep;
 764         unsigned long msiqid, orig_head, head, type, err;
 765
 766         msiqid = (unsigned long) data2;
 767
 768         head = 0xdeadbeef;
 769         err = pci_sun4v_msiq_gethead(pbm->devhandle, msiqid, &head);
 770         if (unlikely(err))
 771                 goto hv_error_get;
 772
 773         if (unlikely(head >= (pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry))))
 774                 goto bad_offset;
 775
 776         head /= sizeof(struct pci_sun4v_msiq_entry);
 777         orig_head = head;
 778         base = (pbm->msi_queues + ((msiqid - pbm->msiq_first) *
 779                                    (pbm->msiq_ent_count *
 780                                     sizeof(struct pci_sun4v_msiq_entry))));
 781         ep = &base[head];
 782         while ((ep->version_type & MSIQ_TYPE_MASK) != 0) {
 783                 type = (ep->version_type & MSIQ_TYPE_MASK) >> MSIQ_TYPE_SHIFT;
 784                 if (unlikely(type != MSIQ_TYPE_MSI32 &&
 785                              type != MSIQ_TYPE_MSI64))
 786                         goto bad_type;
 787
 788                 pci_sun4v_msi_setstate(pbm->devhandle,
 789                                        ep->msi_data /* msi_num */,
 790                                        HV_MSISTATE_IDLE);
 791
 792                 /* Clear the entry.  */
 793                 ep->version_type &= ~MSIQ_TYPE_MASK;
 794
 795                 /* Go to next entry in ring.  */
 796                 head++;
 797                 if (head >= pbm->msiq_ent_count)
 798                         head = 0;
 799                 ep = &base[head];
 800         }
 801
 802         if (likely(head != orig_head)) {
 803                 /* ACK entries by updating head pointer.  */
 804                 head *= sizeof(struct pci_sun4v_msiq_entry);
 805                 err = pci_sun4v_msiq_sethead(pbm->devhandle, msiqid, head);
 806                 if (unlikely(err))
 807                         goto hv_error_set;
 808         }
 809         return;
 810
 811 hv_error_set:
 812         printk(KERN_EMERG "MSI: Hypervisor set head gives error %lu\n", err);
 813         goto hv_error_cont;
 814
 815 hv_error_get:
 816         printk(KERN_EMERG "MSI: Hypervisor get head gives error %lu\n", err);
 817
 818 hv_error_cont:
 819         printk(KERN_EMERG "MSI: devhandle[%x] msiqid[%lx] head[%lu]\n",
 820                pbm->devhandle, msiqid, head);
 821         return;
 822
 823 bad_offset:
 824         printk(KERN_EMERG "MSI: Hypervisor gives bad offset %lx max(%lx)\n",
 825                head, pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry));
 826         return;
 827
 828 bad_type:
 829         printk(KERN_EMERG "MSI: Entry has bad type %lx\n", type);
 830         return;
 831 }
 832
 833 static int msi_bitmap_alloc(struct pci_pbm_info *pbm)
 834 {
 835         unsigned long size, bits_per_ulong;
 836
 837         bits_per_ulong = sizeof(unsigned long) * 8;
 838         size = (pbm->msi_num + (bits_per_ulong - 1)) & ~(bits_per_ulong - 1);
 839         size /= 8;
 840         BUG_ON(size % sizeof(unsigned long));
 841
 842         pbm->msi_bitmap = kzalloc(size, GFP_KERNEL);
 843         if (!pbm->msi_bitmap)
 844                 return -ENOMEM;
 845
 846         return 0;
 847 }
 848
 849 static void msi_bitmap_free(struct pci_pbm_info *pbm)
 850 {
 851         kfree(pbm->msi_bitmap);
 852         pbm->msi_bitmap = NULL;
 853 }
 854
 855 static int msi_queue_alloc(struct pci_pbm_info *pbm)
 856 {
 857         unsigned long q_size, alloc_size, pages, order;
 858         int i;
 859
 860         q_size = pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry);
 861         alloc_size = (pbm->msiq_num * q_size);
 862         order = get_order(alloc_size);
 863         pages = __get_free_pages(GFP_KERNEL | __GFP_COMP, order);
 864         if (pages == 0UL) {
 865                 printk(KERN_ERR "MSI: Cannot allocate MSI queues (o=%lu).\n",
 866                        order);
 867                 return -ENOMEM;
 868         }
 869         memset((char *)pages, 0, PAGE_SIZE << order);
 870         pbm->msi_queues = (void *) pages;
 871
 872         for (i = 0; i < pbm->msiq_num; i++) {
 873                 unsigned long err, base = __pa(pages + (i * q_size));
 874                 unsigned long ret1, ret2;
 875
 876                 err = pci_sun4v_msiq_conf(pbm->devhandle,
 877                                           pbm->msiq_first + i,
 878                                           base, pbm->msiq_ent_count);
 879                 if (err) {
 880                         printk(KERN_ERR "MSI: msiq register fails (err=%lu)\n",
 881                                err);
 882                         goto h_error;
 883                 }
 884
 885                 err = pci_sun4v_msiq_info(pbm->devhandle,
 886                                           pbm->msiq_first + i,
 887                                           &ret1, &ret2);
 888                 if (err) {
 889                         printk(KERN_ERR "MSI: Cannot read msiq (err=%lu)\n",
 890                                err);
 891                         goto h_error;
 892                 }
 893                 if (ret1 != base || ret2 != pbm->msiq_ent_count) {
 894                         printk(KERN_ERR "MSI: Bogus qconf "
 895                                "expected[%lx:%x] got[%lx:%lx]\n",
 896                                base, pbm->msiq_ent_count,
 897                                ret1, ret2);
 898                         goto h_error;
 899                 }
 900         }
 901
 902         return 0;
 903
 904 h_error:
 905         free_pages(pages, order);
 906         return -EINVAL;
 907 }
 908
 909
 910 static int alloc_msi(struct pci_pbm_info *pbm)
 911 {
 912         int i;
 913
 914         for (i = 0; i < pbm->msi_num; i++) {
 915                 if (!test_and_set_bit(i, pbm->msi_bitmap))
 916                         return i + pbm->msi_first;
 917         }
 918
 919         return -ENOENT;
 920 }
 921
 922 static void free_msi(struct pci_pbm_info *pbm, int msi_num)
 923 {
 924         msi_num -= pbm->msi_first;
 925         clear_bit(msi_num, pbm->msi_bitmap);
 926 }
 927
 928 static int pci_sun4v_setup_msi_irq(unsigned int *virt_irq_p,
 929                                    struct pci_dev *pdev,
 930                                    struct msi_desc *entry)
 931 {
 932         struct pci_pbm_info *pbm = pdev->dev.archdata.host_controller;
 933         unsigned long devino, msiqid;
 934         struct msi_msg msg;
 935         int msi_num, err;
 936
 937         *virt_irq_p = 0;
 938
 939         msi_num = alloc_msi(pbm);
 940         if (msi_num < 0)
 941                 return msi_num;
 942
 943         devino = sun4v_build_msi(pbm->devhandle, virt_irq_p,
 944                                  pbm->msiq_first_devino,
 945                                  (pbm->msiq_first_devino +
 946                                   pbm->msiq_num));
 947         err = -ENOMEM;
 948         if (!devino)
 949                 goto out_err;
 950
 951         msiqid = ((devino - pbm->msiq_first_devino) +
 952                   pbm->msiq_first);
 953
 954         err = -EINVAL;
 955         if (pci_sun4v_msiq_setstate(pbm->devhandle, msiqid, HV_MSIQSTATE_IDLE))
 956         if (err)
 957                 goto out_err;
 958
 959         if (pci_sun4v_msiq_setvalid(pbm->devhandle, msiqid, HV_MSIQ_VALID))
 960                 goto out_err;
 961
 962         if (pci_sun4v_msi_setmsiq(pbm->devhandle,
 963                                   msi_num, msiqid,
 964                                   (entry->msi_attrib.is_64 ?
 965                                    HV_MSITYPE_MSI64 : HV_MSITYPE_MSI32)))
 966                 goto out_err;
 967
 968         if (pci_sun4v_msi_setstate(pbm->devhandle, msi_num, HV_MSISTATE_IDLE))
 969                 goto out_err;
 970
 971         if (pci_sun4v_msi_setvalid(pbm->devhandle, msi_num, HV_MSIVALID_VALID))
 972                 goto out_err;
 973
 974         pdev->dev.archdata.msi_num = msi_num;
 975
 976         if (entry->msi_attrib.is_64) {
 977                 msg.address_hi = pbm->msi64_start >> 32;
 978                 msg.address_lo = pbm->msi64_start & 0xffffffff;
 979         } else {
 980                 msg.address_hi = 0;
 981                 msg.address_lo = pbm->msi32_start;
 982         }
 983         msg.data = msi_num;
 984
 985         set_irq_msi(*virt_irq_p, entry);
 986         write_msi_msg(*virt_irq_p, &msg);
 987
 988         irq_install_pre_handler(*virt_irq_p,
 989                                 pci_sun4v_msi_prehandler,
 990                                 pbm, (void *) msiqid);
 991
 992         return 0;
 993
 994 out_err:
 995         free_msi(pbm, msi_num);
 996         sun4v_destroy_msi(*virt_irq_p);
 997         *virt_irq_p = 0;
 998         return err;
 999
1000 }
1001
1002 static void pci_sun4v_teardown_msi_irq(unsigned int virt_irq,
1003                                        struct pci_dev *pdev)
1004 {
1005         struct pci_pbm_info *pbm = pdev->dev.archdata.host_controller;
1006         unsigned long msiqid, err;
1007         unsigned int msi_num;
1008
1009         msi_num = pdev->dev.archdata.msi_num;
1010         err = pci_sun4v_msi_getmsiq(pbm->devhandle, msi_num, &msiqid);
1011         if (err) {
1012                 printk(KERN_ERR "%s: getmsiq gives error %lu\n",
1013                        pbm->name, err);
1014                 return;
1015         }
1016
1017         pci_sun4v_msi_setvalid(pbm->devhandle, msi_num, HV_MSIVALID_INVALID);
1018         pci_sun4v_msiq_setvalid(pbm->devhandle, msiqid, HV_MSIQ_INVALID);
1019
1020         free_msi(pbm, msi_num);
1021
1022         /* The sun4v_destroy_msi() will liberate the devino and thus the MSIQ
1023          * allocation.
1024          */
1025         sun4v_destroy_msi(virt_irq);
1026 }
1027
1028 static void pci_sun4v_msi_init(struct pci_pbm_info *pbm)
1029 {
1030         const u32 *val;
1031         int len;
1032
1033         val = of_get_property(pbm->prom_node, "#msi-eqs", &len);
1034         if (!val || len != 4)
1035                 goto no_msi;
1036         pbm->msiq_num = *val;
1037         if (pbm->msiq_num) {
1038                 const struct msiq_prop {
1039                         u32 first_msiq;
1040                         u32 num_msiq;
1041                         u32 first_devino;
1042                 } *mqp;
1043                 const struct msi_range_prop {
1044                         u32 first_msi;
1045                         u32 num_msi;
1046                 } *mrng;
1047                 const struct addr_range_prop {
1048                         u32 msi32_high;
1049                         u32 msi32_low;
1050                         u32 msi32_len;
1051                         u32 msi64_high;
1052                         u32 msi64_low;
1053                         u32 msi64_len;
1054                 } *arng;
1055
1056                 val = of_get_property(pbm->prom_node, "msi-eq-size", &len);
1057                 if (!val || len != 4)
1058                         goto no_msi;
1059
1060                 pbm->msiq_ent_count = *val;
1061
1062                 mqp = of_get_property(pbm->prom_node,
1063                                       "msi-eq-to-devino", &len);
1064                 if (!mqp || len != sizeof(struct msiq_prop))
1065                         goto no_msi;
1066
1067                 pbm->msiq_first = mqp->first_msiq;
1068                 pbm->msiq_first_devino = mqp->first_devino;
1069
1070                 val = of_get_property(pbm->prom_node, "#msi", &len);
1071                 if (!val || len != 4)
1072                         goto no_msi;
1073                 pbm->msi_num = *val;
1074
1075                 mrng = of_get_property(pbm->prom_node, "msi-ranges", &len);
1076                 if (!mrng || len != sizeof(struct msi_range_prop))
1077                         goto no_msi;
1078                 pbm->msi_first = mrng->first_msi;
1079
1080                 val = of_get_property(pbm->prom_node, "msi-data-mask", &len);
1081                 if (!val || len != 4)
1082                         goto no_msi;
1083                 pbm->msi_data_mask = *val;
1084
1085                 val = of_get_property(pbm->prom_node, "msix-data-width", &len);
1086                 if (!val || len != 4)
1087                         goto no_msi;
1088                 pbm->msix_data_width = *val;
1089
1090                 arng = of_get_property(pbm->prom_node, "msi-address-ranges",
1091                                        &len);
1092                 if (!arng || len != sizeof(struct addr_range_prop))
1093                         goto no_msi;
1094                 pbm->msi32_start = ((u64)arng->msi32_high << 32) |
1095                         (u64) arng->msi32_low;
1096                 pbm->msi64_start = ((u64)arng->msi64_high << 32) |
1097                         (u64) arng->msi64_low;
1098                 pbm->msi32_len = arng->msi32_len;
1099                 pbm->msi64_len = arng->msi64_len;
1100
1101                 if (msi_bitmap_alloc(pbm))
1102                         goto no_msi;
1103
1104                 if (msi_queue_alloc(pbm)) {
1105                         msi_bitmap_free(pbm);
1106                         goto no_msi;
1107                 }
1108
1109                 printk(KERN_INFO "%s: MSI Queue first[%u] num[%u] count[%u] "
1110                        "devino[0x%x]\n",
1111                        pbm->name,
1112                        pbm->msiq_first, pbm->msiq_num,
1113                        pbm->msiq_ent_count,
1114                        pbm->msiq_first_devino);
1115                 printk(KERN_INFO "%s: MSI first[%u] num[%u] mask[0x%x] "
1116                        "width[%u]\n",
1117                        pbm->name,
1118                        pbm->msi_first, pbm->msi_num, pbm->msi_data_mask,
1119                        pbm->msix_data_width);
1120                 printk(KERN_INFO "%s: MSI addr32[0x%lx:0x%x] "
1121                        "addr64[0x%lx:0x%x]\n",
1122                        pbm->name,
1123                        pbm->msi32_start, pbm->msi32_len,
1124                        pbm->msi64_start, pbm->msi64_len);
1125                 printk(KERN_INFO "%s: MSI queues at RA [%p]\n",
1126                        pbm->name,
1127                        pbm->msi_queues);
1128         }
1129         pbm->setup_msi_irq = pci_sun4v_setup_msi_irq;
1130         pbm->teardown_msi_irq = pci_sun4v_teardown_msi_irq;
1131
1132         return;
1133
1134 no_msi:
1135         pbm->msiq_num = 0;
1136         printk(KERN_INFO "%s: No MSI support.\n", pbm->name);
1137 }
1138 #else /* CONFIG_PCI_MSI */
1139 static void pci_sun4v_msi_init(struct pci_pbm_info *pbm)
1140 {
1141 }
1142 #endif /* !(CONFIG_PCI_MSI) */
1143
1144 static void __init pci_sun4v_pbm_init(struct pci_controller_info *p, struct device_node *dp, u32 devhandle)
1145 {
1146         struct pci_pbm_info *pbm;
1147
1148         if (devhandle & 0x40)
1149                 pbm = &p->pbm_B;
1150         else
1151                 pbm = &p->pbm_A;
1152
1153         pbm->next = pci_pbm_root;
1154         pci_pbm_root = pbm;
1155
1156         pbm->scan_bus = pci_sun4v_scan_bus;
1157         pbm->pci_ops = &sun4v_pci_ops;
1158         pbm->config_space_reg_bits = 12;
1159
1160         pbm->index = pci_num_pbms++;
1161
1162         pbm->parent = p;
1163         pbm->prom_node = dp;
1164
1165         pbm->devhandle = devhandle;
1166
1167         pbm->name = dp->full_name;
1168
1169         printk("%s: SUN4V PCI Bus Module\n", pbm->name);
1170
1171         pci_determine_mem_io_space(pbm);
1172
1173         pci_get_pbm_props(pbm);
1174         pci_sun4v_iommu_init(pbm);
1175         pci_sun4v_msi_init(pbm);
1176 }
1177
1178 void __init sun4v_pci_init(struct device_node *dp, char *model_name)
1179 {
1180         static int hvapi_negotiated = 0;
1181         struct pci_controller_info *p;
1182         struct pci_pbm_info *pbm;
1183         struct iommu *iommu;
1184         struct property *prop;
1185         struct linux_prom64_registers *regs;
1186         u32 devhandle;
1187         int i;
1188
1189         if (!hvapi_negotiated++) {
1190                 int err = sun4v_hvapi_register(HV_GRP_PCI,
1191                                                vpci_major,
1192                                                &vpci_minor);
1193
1194                 if (err) {
1195                         prom_printf("SUN4V_PCI: Could not register hvapi, "
1196                                     "err=%d\n", err);
1197                         prom_halt();
1198                 }
1199                 printk("SUN4V_PCI: Registered hvapi major[%lu] minor[%lu]\n",
1200                        vpci_major, vpci_minor);
1201
1202                 dma_ops = &sun4v_dma_ops;
1203         }
1204
1205         prop = of_find_property(dp, "reg", NULL);
1206         regs = prop->value;
1207
1208         devhandle = (regs->phys_addr >> 32UL) & 0x0fffffff;
1209
1210         for (pbm = pci_pbm_root; pbm; pbm = pbm->next) {
1211                 if (pbm->devhandle == (devhandle ^ 0x40)) {
1212                         pci_sun4v_pbm_init(pbm->parent, dp, devhandle);
1213                         return;
1214                 }
1215         }
1216
1217         for_each_possible_cpu(i) {
1218                 unsigned long page = get_zeroed_page(GFP_ATOMIC);
1219
1220                 if (!page)
1221                         goto fatal_memory_error;
1222
1223                 per_cpu(iommu_batch, i).pglist = (u64 *) page;
1224         }
1225
1226         p = kzalloc(sizeof(struct pci_controller_info), GFP_ATOMIC);
1227         if (!p)
1228                 goto fatal_memory_error;
1229
1230         iommu = kzalloc(sizeof(struct iommu), GFP_ATOMIC);
1231         if (!iommu)
1232                 goto fatal_memory_error;
1233
1234         p->pbm_A.iommu = iommu;
1235
1236         iommu = kzalloc(sizeof(struct iommu), GFP_ATOMIC);
1237         if (!iommu)
1238                 goto fatal_memory_error;
1239
1240         p->pbm_B.iommu = iommu;
1241
1242         /* Like PSYCHO and SCHIZO we have a 2GB aligned area
1243          * for memory space.
1244          */
1245         pci_memspace_mask = 0x7fffffffUL;
1246
1247         pci_sun4v_pbm_init(p, dp, devhandle);
1248         return;
1249
1250 fatal_memory_error:
1251         prom_printf("SUN4V_PCI: Fatal memory allocation error.\n");
1252         prom_halt();
1253 }