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 pci_dev  *pdev;          /* 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, pci_iommu_batch);
  44
  45 /* Interrupts must be disabled.  */
  46 static inline void pci_iommu_batch_start(struct pci_dev *pdev, unsigned long prot, unsigned long entry)
  47 {
  48         struct iommu_batch *p = &__get_cpu_var(pci_iommu_batch);
  49
  50         p->pdev         = pdev;
  51         p->prot         = prot;
  52         p->entry        = entry;
  53         p->npages       = 0;
  54 }
  55
  56 /* Interrupts must be disabled.  */
  57 static long pci_iommu_batch_flush(struct iommu_batch *p)
  58 {
  59         struct pci_pbm_info *pbm = p->pdev->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("pci_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 pci_iommu_batch_add(u64 phys_page)
  94 {
  95         struct iommu_batch *p = &__get_cpu_var(pci_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 pci_iommu_batch_flush(p);
 102
 103         return 0;
 104 }
 105
 106 /* Interrupts must be disabled.  */
 107 static inline long pci_iommu_batch_end(void)
 108 {
 109         struct iommu_batch *p = &__get_cpu_var(pci_iommu_batch);
 110
 111         BUG_ON(p->npages >= PGLIST_NENTS);
 112
 113         return pci_iommu_batch_flush(p);
 114 }
 115
 116 static long pci_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 pci_arena_free(struct iommu_arena *arena, unsigned long base, unsigned long npages)
 156 {
 157         unsigned long i;
 158
 159         for (i = base; i < (base + npages); i++)
 160                 __clear_bit(i, arena->map);
 161 }
 162
 163 static void *pci_4v_alloc_consistent(struct pci_dev *pdev, size_t size, dma_addr_t *dma_addrp, gfp_t gfp)
 164 {
 165         struct iommu *iommu;
 166         unsigned long flags, order, first_page, npages, n;
 167         void *ret;
 168         long entry;
 169
 170         size = IO_PAGE_ALIGN(size);
 171         order = get_order(size);
 172         if (unlikely(order >= MAX_ORDER))
 173                 return NULL;
 174
 175         npages = size >> IO_PAGE_SHIFT;
 176
 177         first_page = __get_free_pages(gfp, order);
 178         if (unlikely(first_page == 0UL))
 179                 return NULL;
 180
 181         memset((char *)first_page, 0, PAGE_SIZE << order);
 182
 183         iommu = pdev->dev.archdata.iommu;
 184
 185         spin_lock_irqsave(&iommu->lock, flags);
 186         entry = pci_arena_alloc(&iommu->arena, npages);
 187         spin_unlock_irqrestore(&iommu->lock, flags);
 188
 189         if (unlikely(entry < 0L))
 190                 goto arena_alloc_fail;
 191
 192         *dma_addrp = (iommu->page_table_map_base +
 193                       (entry << IO_PAGE_SHIFT));
 194         ret = (void *) first_page;
 195         first_page = __pa(first_page);
 196
 197         local_irq_save(flags);
 198
 199         pci_iommu_batch_start(pdev,
 200                               (HV_PCI_MAP_ATTR_READ |
 201                                HV_PCI_MAP_ATTR_WRITE),
 202                               entry);
 203
 204         for (n = 0; n < npages; n++) {
 205                 long err = pci_iommu_batch_add(first_page + (n * PAGE_SIZE));
 206                 if (unlikely(err < 0L))
 207                         goto iommu_map_fail;
 208         }
 209
 210         if (unlikely(pci_iommu_batch_end() < 0L))
 211                 goto iommu_map_fail;
 212
 213         local_irq_restore(flags);
 214
 215         return ret;
 216
 217 iommu_map_fail:
 218         /* Interrupts are disabled.  */
 219         spin_lock(&iommu->lock);
 220         pci_arena_free(&iommu->arena, entry, npages);
 221         spin_unlock_irqrestore(&iommu->lock, flags);
 222
 223 arena_alloc_fail:
 224         free_pages(first_page, order);
 225         return NULL;
 226 }
 227
 228 static void pci_4v_free_consistent(struct pci_dev *pdev, size_t size, void *cpu, dma_addr_t dvma)
 229 {
 230         struct pci_pbm_info *pbm;
 231         struct iommu *iommu;
 232         unsigned long flags, order, npages, entry;
 233         u32 devhandle;
 234
 235         npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT;
 236         iommu = pdev->dev.archdata.iommu;
 237         pbm = pdev->dev.archdata.host_controller;
 238         devhandle = pbm->devhandle;
 239         entry = ((dvma - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
 240
 241         spin_lock_irqsave(&iommu->lock, flags);
 242
 243         pci_arena_free(&iommu->arena, entry, npages);
 244
 245         do {
 246                 unsigned long num;
 247
 248                 num = pci_sun4v_iommu_demap(devhandle, HV_PCI_TSBID(0, entry),
 249                                             npages);
 250                 entry += num;
 251                 npages -= num;
 252         } while (npages != 0);
 253
 254         spin_unlock_irqrestore(&iommu->lock, flags);
 255
 256         order = get_order(size);
 257         if (order < 10)
 258                 free_pages((unsigned long)cpu, order);
 259 }
 260
 261 static dma_addr_t pci_4v_map_single(struct pci_dev *pdev, void *ptr, size_t sz, int direction)
 262 {
 263         struct iommu *iommu;
 264         unsigned long flags, npages, oaddr;
 265         unsigned long i, base_paddr;
 266         u32 bus_addr, ret;
 267         unsigned long prot;
 268         long entry;
 269
 270         iommu = pdev->dev.archdata.iommu;
 271
 272         if (unlikely(direction == PCI_DMA_NONE))
 273                 goto bad;
 274
 275         oaddr = (unsigned long)ptr;
 276         npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK);
 277         npages >>= IO_PAGE_SHIFT;
 278
 279         spin_lock_irqsave(&iommu->lock, flags);
 280         entry = pci_arena_alloc(&iommu->arena, npages);
 281         spin_unlock_irqrestore(&iommu->lock, flags);
 282
 283         if (unlikely(entry < 0L))
 284                 goto bad;
 285
 286         bus_addr = (iommu->page_table_map_base +
 287                     (entry << IO_PAGE_SHIFT));
 288         ret = bus_addr | (oaddr & ~IO_PAGE_MASK);
 289         base_paddr = __pa(oaddr & IO_PAGE_MASK);
 290         prot = HV_PCI_MAP_ATTR_READ;
 291         if (direction != PCI_DMA_TODEVICE)
 292                 prot |= HV_PCI_MAP_ATTR_WRITE;
 293
 294         local_irq_save(flags);
 295
 296         pci_iommu_batch_start(pdev, prot, entry);
 297
 298         for (i = 0; i < npages; i++, base_paddr += IO_PAGE_SIZE) {
 299                 long err = pci_iommu_batch_add(base_paddr);
 300                 if (unlikely(err < 0L))
 301                         goto iommu_map_fail;
 302         }
 303         if (unlikely(pci_iommu_batch_end() < 0L))
 304                 goto iommu_map_fail;
 305
 306         local_irq_restore(flags);
 307
 308         return ret;
 309
 310 bad:
 311         if (printk_ratelimit())
 312                 WARN_ON(1);
 313         return PCI_DMA_ERROR_CODE;
 314
 315 iommu_map_fail:
 316         /* Interrupts are disabled.  */
 317         spin_lock(&iommu->lock);
 318         pci_arena_free(&iommu->arena, entry, npages);
 319         spin_unlock_irqrestore(&iommu->lock, flags);
 320
 321         return PCI_DMA_ERROR_CODE;
 322 }
 323
 324 static void pci_4v_unmap_single(struct pci_dev *pdev, dma_addr_t bus_addr, size_t sz, int direction)
 325 {
 326         struct pci_pbm_info *pbm;
 327         struct iommu *iommu;
 328         unsigned long flags, npages;
 329         long entry;
 330         u32 devhandle;
 331
 332         if (unlikely(direction == PCI_DMA_NONE)) {
 333                 if (printk_ratelimit())
 334                         WARN_ON(1);
 335                 return;
 336         }
 337
 338         iommu = pdev->dev.archdata.iommu;
 339         pbm = pdev->dev.archdata.host_controller;
 340         devhandle = pbm->devhandle;
 341
 342         npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK);
 343         npages >>= IO_PAGE_SHIFT;
 344         bus_addr &= IO_PAGE_MASK;
 345
 346         spin_lock_irqsave(&iommu->lock, flags);
 347
 348         entry = (bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT;
 349         pci_arena_free(&iommu->arena, entry, npages);
 350
 351         do {
 352                 unsigned long num;
 353
 354                 num = pci_sun4v_iommu_demap(devhandle, HV_PCI_TSBID(0, entry),
 355                                             npages);
 356                 entry += num;
 357                 npages -= num;
 358         } while (npages != 0);
 359
 360         spin_unlock_irqrestore(&iommu->lock, flags);
 361 }
 362
 363 #define SG_ENT_PHYS_ADDRESS(SG) \
 364         (__pa(page_address((SG)->page)) + (SG)->offset)
 365
 366 static inline long fill_sg(long entry, struct pci_dev *pdev,
 367                            struct scatterlist *sg,
 368                            int nused, int nelems, unsigned long prot)
 369 {
 370         struct scatterlist *dma_sg = sg;
 371         struct scatterlist *sg_end = sg + nelems;
 372         unsigned long flags;
 373         int i;
 374
 375         local_irq_save(flags);
 376
 377         pci_iommu_batch_start(pdev, prot, entry);
 378
 379         for (i = 0; i < nused; i++) {
 380                 unsigned long pteval = ~0UL;
 381                 u32 dma_npages;
 382
 383                 dma_npages = ((dma_sg->dma_address & (IO_PAGE_SIZE - 1UL)) +
 384                               dma_sg->dma_length +
 385                               ((IO_PAGE_SIZE - 1UL))) >> IO_PAGE_SHIFT;
 386                 do {
 387                         unsigned long offset;
 388                         signed int len;
 389
 390                         /* If we are here, we know we have at least one
 391                          * more page to map.  So walk forward until we
 392                          * hit a page crossing, and begin creating new
 393                          * mappings from that spot.
 394                          */
 395                         for (;;) {
 396                                 unsigned long tmp;
 397
 398                                 tmp = SG_ENT_PHYS_ADDRESS(sg);
 399                                 len = sg->length;
 400                                 if (((tmp ^ pteval) >> IO_PAGE_SHIFT) != 0UL) {
 401                                         pteval = tmp & IO_PAGE_MASK;
 402                                         offset = tmp & (IO_PAGE_SIZE - 1UL);
 403                                         break;
 404                                 }
 405                                 if (((tmp ^ (tmp + len - 1UL)) >> IO_PAGE_SHIFT) != 0UL) {
 406                                         pteval = (tmp + IO_PAGE_SIZE) & IO_PAGE_MASK;
 407                                         offset = 0UL;
 408                                         len -= (IO_PAGE_SIZE - (tmp & (IO_PAGE_SIZE - 1UL)));
 409                                         break;
 410                                 }
 411                                 sg++;
 412                         }
 413
 414                         pteval = (pteval & IOPTE_PAGE);
 415                         while (len > 0) {
 416                                 long err;
 417
 418                                 err = pci_iommu_batch_add(pteval);
 419                                 if (unlikely(err < 0L))
 420                                         goto iommu_map_failed;
 421
 422                                 pteval += IO_PAGE_SIZE;
 423                                 len -= (IO_PAGE_SIZE - offset);
 424                                 offset = 0;
 425                                 dma_npages--;
 426                         }
 427
 428                         pteval = (pteval & IOPTE_PAGE) + len;
 429                         sg++;
 430
 431                         /* Skip over any tail mappings we've fully mapped,
 432                          * adjusting pteval along the way.  Stop when we
 433                          * detect a page crossing event.
 434                          */
 435                         while (sg < sg_end &&
 436                                (pteval << (64 - IO_PAGE_SHIFT)) != 0UL &&
 437                                (pteval == SG_ENT_PHYS_ADDRESS(sg)) &&
 438                                ((pteval ^
 439                                  (SG_ENT_PHYS_ADDRESS(sg) + sg->length - 1UL)) >> IO_PAGE_SHIFT) == 0UL) {
 440                                 pteval += sg->length;
 441                                 sg++;
 442                         }
 443                         if ((pteval << (64 - IO_PAGE_SHIFT)) == 0UL)
 444                                 pteval = ~0UL;
 445                 } while (dma_npages != 0);
 446                 dma_sg++;
 447         }
 448
 449         if (unlikely(pci_iommu_batch_end() < 0L))
 450                 goto iommu_map_failed;
 451
 452         local_irq_restore(flags);
 453         return 0;
 454
 455 iommu_map_failed:
 456         local_irq_restore(flags);
 457         return -1L;
 458 }
 459
 460 static int pci_4v_map_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int direction)
 461 {
 462         struct iommu *iommu;
 463         unsigned long flags, npages, prot;
 464         u32 dma_base;
 465         struct scatterlist *sgtmp;
 466         long entry, err;
 467         int used;
 468
 469         /* Fast path single entry scatterlists. */
 470         if (nelems == 1) {
 471                 sglist->dma_address =
 472                         pci_4v_map_single(pdev,
 473                                           (page_address(sglist->page) + sglist->offset),
 474                                           sglist->length, direction);
 475                 if (unlikely(sglist->dma_address == PCI_DMA_ERROR_CODE))
 476                         return 0;
 477                 sglist->dma_length = sglist->length;
 478                 return 1;
 479         }
 480
 481         iommu = pdev->dev.archdata.iommu;
 482
 483         if (unlikely(direction == PCI_DMA_NONE))
 484                 goto bad;
 485
 486         /* Step 1: Prepare scatter list. */
 487         npages = prepare_sg(sglist, nelems);
 488
 489         /* Step 2: Allocate a cluster and context, if necessary. */
 490         spin_lock_irqsave(&iommu->lock, flags);
 491         entry = pci_arena_alloc(&iommu->arena, npages);
 492         spin_unlock_irqrestore(&iommu->lock, flags);
 493
 494         if (unlikely(entry < 0L))
 495                 goto bad;
 496
 497         dma_base = iommu->page_table_map_base +
 498                 (entry << IO_PAGE_SHIFT);
 499
 500         /* Step 3: Normalize DMA addresses. */
 501         used = nelems;
 502
 503         sgtmp = sglist;
 504         while (used && sgtmp->dma_length) {
 505                 sgtmp->dma_address += dma_base;
 506                 sgtmp++;
 507                 used--;
 508         }
 509         used = nelems - used;
 510
 511         /* Step 4: Create the mappings. */
 512         prot = HV_PCI_MAP_ATTR_READ;
 513         if (direction != PCI_DMA_TODEVICE)
 514                 prot |= HV_PCI_MAP_ATTR_WRITE;
 515
 516         err = fill_sg(entry, pdev, sglist, used, nelems, prot);
 517         if (unlikely(err < 0L))
 518                 goto iommu_map_failed;
 519
 520         return used;
 521
 522 bad:
 523         if (printk_ratelimit())
 524                 WARN_ON(1);
 525         return 0;
 526
 527 iommu_map_failed:
 528         spin_lock_irqsave(&iommu->lock, flags);
 529         pci_arena_free(&iommu->arena, entry, npages);
 530         spin_unlock_irqrestore(&iommu->lock, flags);
 531
 532         return 0;
 533 }
 534
 535 static void pci_4v_unmap_sg(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int direction)
 536 {
 537         struct pci_pbm_info *pbm;
 538         struct iommu *iommu;
 539         unsigned long flags, i, npages;
 540         long entry;
 541         u32 devhandle, bus_addr;
 542
 543         if (unlikely(direction == PCI_DMA_NONE)) {
 544                 if (printk_ratelimit())
 545                         WARN_ON(1);
 546         }
 547
 548         iommu = pdev->dev.archdata.iommu;
 549         pbm = pdev->dev.archdata.host_controller;
 550         devhandle = pbm->devhandle;
 551
 552         bus_addr = sglist->dma_address & IO_PAGE_MASK;
 553
 554         for (i = 1; i < nelems; i++)
 555                 if (sglist[i].dma_length == 0)
 556                         break;
 557         i--;
 558         npages = (IO_PAGE_ALIGN(sglist[i].dma_address + sglist[i].dma_length) -
 559                   bus_addr) >> IO_PAGE_SHIFT;
 560
 561         entry = ((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
 562
 563         spin_lock_irqsave(&iommu->lock, flags);
 564
 565         pci_arena_free(&iommu->arena, entry, npages);
 566
 567         do {
 568                 unsigned long num;
 569
 570                 num = pci_sun4v_iommu_demap(devhandle, HV_PCI_TSBID(0, entry),
 571                                             npages);
 572                 entry += num;
 573                 npages -= num;
 574         } while (npages != 0);
 575
 576         spin_unlock_irqrestore(&iommu->lock, flags);
 577 }
 578
 579 static void pci_4v_dma_sync_single_for_cpu(struct pci_dev *pdev, dma_addr_t bus_addr, size_t sz, int direction)
 580 {
 581         /* Nothing to do... */
 582 }
 583
 584 static void pci_4v_dma_sync_sg_for_cpu(struct pci_dev *pdev, struct scatterlist *sglist, int nelems, int direction)
 585 {
 586         /* Nothing to do... */
 587 }
 588
 589 const struct pci_iommu_ops pci_sun4v_iommu_ops = {
 590         .alloc_consistent               = pci_4v_alloc_consistent,
 591         .free_consistent                = pci_4v_free_consistent,
 592         .map_single                     = pci_4v_map_single,
 593         .unmap_single                   = pci_4v_unmap_single,
 594         .map_sg                         = pci_4v_map_sg,
 595         .unmap_sg                       = pci_4v_unmap_sg,
 596         .dma_sync_single_for_cpu        = pci_4v_dma_sync_single_for_cpu,
 597         .dma_sync_sg_for_cpu            = pci_4v_dma_sync_sg_for_cpu,
 598 };
 599
 600 static void pci_sun4v_scan_bus(struct pci_pbm_info *pbm)
 601 {
 602         struct property *prop;
 603         struct device_node *dp;
 604
 605         dp = pbm->prom_node;
 606         prop = of_find_property(dp, "66mhz-capable", NULL);
 607         pbm->is_66mhz_capable = (prop != NULL);
 608         pbm->pci_bus = pci_scan_one_pbm(pbm);
 609
 610         /* XXX register error interrupt handlers XXX */
 611 }
 612
 613 static unsigned long probe_existing_entries(struct pci_pbm_info *pbm,
 614                                             struct iommu *iommu)
 615 {
 616         struct iommu_arena *arena = &iommu->arena;
 617         unsigned long i, cnt = 0;
 618         u32 devhandle;
 619
 620         devhandle = pbm->devhandle;
 621         for (i = 0; i < arena->limit; i++) {
 622                 unsigned long ret, io_attrs, ra;
 623
 624                 ret = pci_sun4v_iommu_getmap(devhandle,
 625                                              HV_PCI_TSBID(0, i),
 626                                              &io_attrs, &ra);
 627                 if (ret == HV_EOK) {
 628                         if (page_in_phys_avail(ra)) {
 629                                 pci_sun4v_iommu_demap(devhandle,
 630                                                       HV_PCI_TSBID(0, i), 1);
 631                         } else {
 632                                 cnt++;
 633                                 __set_bit(i, arena->map);
 634                         }
 635                 }
 636         }
 637
 638         return cnt;
 639 }
 640
 641 static void pci_sun4v_iommu_init(struct pci_pbm_info *pbm)
 642 {
 643         struct iommu *iommu = pbm->iommu;
 644         struct property *prop;
 645         unsigned long num_tsb_entries, sz, tsbsize;
 646         u32 vdma[2], dma_mask, dma_offset;
 647
 648         prop = of_find_property(pbm->prom_node, "virtual-dma", NULL);
 649         if (prop) {
 650                 u32 *val = prop->value;
 651
 652                 vdma[0] = val[0];
 653                 vdma[1] = val[1];
 654         } else {
 655                 /* No property, use default values. */
 656                 vdma[0] = 0x80000000;
 657                 vdma[1] = 0x80000000;
 658         }
 659
 660         if ((vdma[0] | vdma[1]) & ~IO_PAGE_MASK) {
 661                 prom_printf("PCI-SUN4V: strange virtual-dma[%08x:%08x].\n",
 662                             vdma[0], vdma[1]);
 663                 prom_halt();
 664         };
 665
 666         dma_mask = (roundup_pow_of_two(vdma[1]) - 1UL);
 667         num_tsb_entries = vdma[1] / IO_PAGE_SIZE;
 668         tsbsize = num_tsb_entries * sizeof(iopte_t);
 669
 670         dma_offset = vdma[0];
 671
 672         /* Setup initial software IOMMU state. */
 673         spin_lock_init(&iommu->lock);
 674         iommu->ctx_lowest_free = 1;
 675         iommu->page_table_map_base = dma_offset;
 676         iommu->dma_addr_mask = dma_mask;
 677
 678         /* Allocate and initialize the free area map.  */
 679         sz = (num_tsb_entries + 7) / 8;
 680         sz = (sz + 7UL) & ~7UL;
 681         iommu->arena.map = kzalloc(sz, GFP_KERNEL);
 682         if (!iommu->arena.map) {
 683                 prom_printf("PCI_IOMMU: Error, kmalloc(arena.map) failed.\n");
 684                 prom_halt();
 685         }
 686         iommu->arena.limit = num_tsb_entries;
 687
 688         sz = probe_existing_entries(pbm, iommu);
 689         if (sz)
 690                 printk("%s: Imported %lu TSB entries from OBP\n",
 691                        pbm->name, sz);
 692 }
 693
 694 #ifdef CONFIG_PCI_MSI
 695 struct pci_sun4v_msiq_entry {
 696         u64             version_type;
 697 #define MSIQ_VERSION_MASK               0xffffffff00000000UL
 698 #define MSIQ_VERSION_SHIFT              32
 699 #define MSIQ_TYPE_MASK                  0x00000000000000ffUL
 700 #define MSIQ_TYPE_SHIFT                 0
 701 #define MSIQ_TYPE_NONE                  0x00
 702 #define MSIQ_TYPE_MSG                   0x01
 703 #define MSIQ_TYPE_MSI32                 0x02
 704 #define MSIQ_TYPE_MSI64                 0x03
 705 #define MSIQ_TYPE_INTX                  0x08
 706 #define MSIQ_TYPE_NONE2                 0xff
 707
 708         u64             intx_sysino;
 709         u64             reserved1;
 710         u64             stick;
 711         u64             req_id;  /* bus/device/func */
 712 #define MSIQ_REQID_BUS_MASK             0xff00UL
 713 #define MSIQ_REQID_BUS_SHIFT            8
 714 #define MSIQ_REQID_DEVICE_MASK          0x00f8UL
 715 #define MSIQ_REQID_DEVICE_SHIFT         3
 716 #define MSIQ_REQID_FUNC_MASK            0x0007UL
 717 #define MSIQ_REQID_FUNC_SHIFT           0
 718
 719         u64             msi_address;
 720
 721         /* The format of this value is message type dependent.
 722          * For MSI bits 15:0 are the data from the MSI packet.
 723          * For MSI-X bits 31:0 are the data from the MSI packet.
 724          * For MSG, the message code and message routing code where:
 725          *      bits 39:32 is the bus/device/fn of the msg target-id
 726          *      bits 18:16 is the message routing code
 727          *      bits 7:0 is the message code
 728          * For INTx the low order 2-bits are:
 729          *      00 - INTA
 730          *      01 - INTB
 731          *      10 - INTC
 732          *      11 - INTD
 733          */
 734         u64             msi_data;
 735
 736         u64             reserved2;
 737 };
 738
 739 /* For now this just runs as a pre-handler for the real interrupt handler.
 740  * So we just walk through the queue and ACK all the entries, update the
 741  * head pointer, and return.
 742  *
 743  * In the longer term it would be nice to do something more integrated
 744  * wherein we can pass in some of this MSI info to the drivers.  This
 745  * would be most useful for PCIe fabric error messages, although we could
 746  * invoke those directly from the loop here in order to pass the info around.
 747  */
 748 static void pci_sun4v_msi_prehandler(unsigned int ino, void *data1, void *data2)
 749 {
 750         struct pci_pbm_info *pbm = data1;
 751         struct pci_sun4v_msiq_entry *base, *ep;
 752         unsigned long msiqid, orig_head, head, type, err;
 753
 754         msiqid = (unsigned long) data2;
 755
 756         head = 0xdeadbeef;
 757         err = pci_sun4v_msiq_gethead(pbm->devhandle, msiqid, &head);
 758         if (unlikely(err))
 759                 goto hv_error_get;
 760
 761         if (unlikely(head >= (pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry))))
 762                 goto bad_offset;
 763
 764         head /= sizeof(struct pci_sun4v_msiq_entry);
 765         orig_head = head;
 766         base = (pbm->msi_queues + ((msiqid - pbm->msiq_first) *
 767                                    (pbm->msiq_ent_count *
 768                                     sizeof(struct pci_sun4v_msiq_entry))));
 769         ep = &base[head];
 770         while ((ep->version_type & MSIQ_TYPE_MASK) != 0) {
 771                 type = (ep->version_type & MSIQ_TYPE_MASK) >> MSIQ_TYPE_SHIFT;
 772                 if (unlikely(type != MSIQ_TYPE_MSI32 &&
 773                              type != MSIQ_TYPE_MSI64))
 774                         goto bad_type;
 775
 776                 pci_sun4v_msi_setstate(pbm->devhandle,
 777                                        ep->msi_data /* msi_num */,
 778                                        HV_MSISTATE_IDLE);
 779
 780                 /* Clear the entry.  */
 781                 ep->version_type &= ~MSIQ_TYPE_MASK;
 782
 783                 /* Go to next entry in ring.  */
 784                 head++;
 785                 if (head >= pbm->msiq_ent_count)
 786                         head = 0;
 787                 ep = &base[head];
 788         }
 789
 790         if (likely(head != orig_head)) {
 791                 /* ACK entries by updating head pointer.  */
 792                 head *= sizeof(struct pci_sun4v_msiq_entry);
 793                 err = pci_sun4v_msiq_sethead(pbm->devhandle, msiqid, head);
 794                 if (unlikely(err))
 795                         goto hv_error_set;
 796         }
 797         return;
 798
 799 hv_error_set:
 800         printk(KERN_EMERG "MSI: Hypervisor set head gives error %lu\n", err);
 801         goto hv_error_cont;
 802
 803 hv_error_get:
 804         printk(KERN_EMERG "MSI: Hypervisor get head gives error %lu\n", err);
 805
 806 hv_error_cont:
 807         printk(KERN_EMERG "MSI: devhandle[%x] msiqid[%lx] head[%lu]\n",
 808                pbm->devhandle, msiqid, head);
 809         return;
 810
 811 bad_offset:
 812         printk(KERN_EMERG "MSI: Hypervisor gives bad offset %lx max(%lx)\n",
 813                head, pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry));
 814         return;
 815
 816 bad_type:
 817         printk(KERN_EMERG "MSI: Entry has bad type %lx\n", type);
 818         return;
 819 }
 820
 821 static int msi_bitmap_alloc(struct pci_pbm_info *pbm)
 822 {
 823         unsigned long size, bits_per_ulong;
 824
 825         bits_per_ulong = sizeof(unsigned long) * 8;
 826         size = (pbm->msi_num + (bits_per_ulong - 1)) & ~(bits_per_ulong - 1);
 827         size /= 8;
 828         BUG_ON(size % sizeof(unsigned long));
 829
 830         pbm->msi_bitmap = kzalloc(size, GFP_KERNEL);
 831         if (!pbm->msi_bitmap)
 832                 return -ENOMEM;
 833
 834         return 0;
 835 }
 836
 837 static void msi_bitmap_free(struct pci_pbm_info *pbm)
 838 {
 839         kfree(pbm->msi_bitmap);
 840         pbm->msi_bitmap = NULL;
 841 }
 842
 843 static int msi_queue_alloc(struct pci_pbm_info *pbm)
 844 {
 845         unsigned long q_size, alloc_size, pages, order;
 846         int i;
 847
 848         q_size = pbm->msiq_ent_count * sizeof(struct pci_sun4v_msiq_entry);
 849         alloc_size = (pbm->msiq_num * q_size);
 850         order = get_order(alloc_size);
 851         pages = __get_free_pages(GFP_KERNEL | __GFP_COMP, order);
 852         if (pages == 0UL) {
 853                 printk(KERN_ERR "MSI: Cannot allocate MSI queues (o=%lu).\n",
 854                        order);
 855                 return -ENOMEM;
 856         }
 857         memset((char *)pages, 0, PAGE_SIZE << order);
 858         pbm->msi_queues = (void *) pages;
 859
 860         for (i = 0; i < pbm->msiq_num; i++) {
 861                 unsigned long err, base = __pa(pages + (i * q_size));
 862                 unsigned long ret1, ret2;
 863
 864                 err = pci_sun4v_msiq_conf(pbm->devhandle,
 865                                           pbm->msiq_first + i,
 866                                           base, pbm->msiq_ent_count);
 867                 if (err) {
 868                         printk(KERN_ERR "MSI: msiq register fails (err=%lu)\n",
 869                                err);
 870                         goto h_error;
 871                 }
 872
 873                 err = pci_sun4v_msiq_info(pbm->devhandle,
 874                                           pbm->msiq_first + i,
 875                                           &ret1, &ret2);
 876                 if (err) {
 877                         printk(KERN_ERR "MSI: Cannot read msiq (err=%lu)\n",
 878                                err);
 879                         goto h_error;
 880                 }
 881                 if (ret1 != base || ret2 != pbm->msiq_ent_count) {
 882                         printk(KERN_ERR "MSI: Bogus qconf "
 883                                "expected[%lx:%x] got[%lx:%lx]\n",
 884                                base, pbm->msiq_ent_count,
 885                                ret1, ret2);
 886                         goto h_error;
 887                 }
 888         }
 889
 890         return 0;
 891
 892 h_error:
 893         free_pages(pages, order);
 894         return -EINVAL;
 895 }
 896
 897
 898 static int alloc_msi(struct pci_pbm_info *pbm)
 899 {
 900         int i;
 901
 902         for (i = 0; i < pbm->msi_num; i++) {
 903                 if (!test_and_set_bit(i, pbm->msi_bitmap))
 904                         return i + pbm->msi_first;
 905         }
 906
 907         return -ENOENT;
 908 }
 909
 910 static void free_msi(struct pci_pbm_info *pbm, int msi_num)
 911 {
 912         msi_num -= pbm->msi_first;
 913         clear_bit(msi_num, pbm->msi_bitmap);
 914 }
 915
 916 static int pci_sun4v_setup_msi_irq(unsigned int *virt_irq_p,
 917                                    struct pci_dev *pdev,
 918                                    struct msi_desc *entry)
 919 {
 920         struct pci_pbm_info *pbm = pdev->dev.archdata.host_controller;
 921         unsigned long devino, msiqid;
 922         struct msi_msg msg;
 923         int msi_num, err;
 924
 925         *virt_irq_p = 0;
 926
 927         msi_num = alloc_msi(pbm);
 928         if (msi_num < 0)
 929                 return msi_num;
 930
 931         devino = sun4v_build_msi(pbm->devhandle, virt_irq_p,
 932                                  pbm->msiq_first_devino,
 933                                  (pbm->msiq_first_devino +
 934                                   pbm->msiq_num));
 935         err = -ENOMEM;
 936         if (!devino)
 937                 goto out_err;
 938
 939         msiqid = ((devino - pbm->msiq_first_devino) +
 940                   pbm->msiq_first);
 941
 942         err = -EINVAL;
 943         if (pci_sun4v_msiq_setstate(pbm->devhandle, msiqid, HV_MSIQSTATE_IDLE))
 944         if (err)
 945                 goto out_err;
 946
 947         if (pci_sun4v_msiq_setvalid(pbm->devhandle, msiqid, HV_MSIQ_VALID))
 948                 goto out_err;
 949
 950         if (pci_sun4v_msi_setmsiq(pbm->devhandle,
 951                                   msi_num, msiqid,
 952                                   (entry->msi_attrib.is_64 ?
 953                                    HV_MSITYPE_MSI64 : HV_MSITYPE_MSI32)))
 954                 goto out_err;
 955
 956         if (pci_sun4v_msi_setstate(pbm->devhandle, msi_num, HV_MSISTATE_IDLE))
 957                 goto out_err;
 958
 959         if (pci_sun4v_msi_setvalid(pbm->devhandle, msi_num, HV_MSIVALID_VALID))
 960                 goto out_err;
 961
 962         pdev->dev.archdata.msi_num = msi_num;
 963
 964         if (entry->msi_attrib.is_64) {
 965                 msg.address_hi = pbm->msi64_start >> 32;
 966                 msg.address_lo = pbm->msi64_start & 0xffffffff;
 967         } else {
 968                 msg.address_hi = 0;
 969                 msg.address_lo = pbm->msi32_start;
 970         }
 971         msg.data = msi_num;
 972
 973         set_irq_msi(*virt_irq_p, entry);
 974         write_msi_msg(*virt_irq_p, &msg);
 975
 976         irq_install_pre_handler(*virt_irq_p,
 977                                 pci_sun4v_msi_prehandler,
 978                                 pbm, (void *) msiqid);
 979
 980         return 0;
 981
 982 out_err:
 983         free_msi(pbm, msi_num);
 984         sun4v_destroy_msi(*virt_irq_p);
 985         *virt_irq_p = 0;
 986         return err;
 987
 988 }
 989
 990 static void pci_sun4v_teardown_msi_irq(unsigned int virt_irq,
 991                                        struct pci_dev *pdev)
 992 {
 993         struct pci_pbm_info *pbm = pdev->dev.archdata.host_controller;
 994         unsigned long msiqid, err;
 995         unsigned int msi_num;
 996
 997         msi_num = pdev->dev.archdata.msi_num;
 998         err = pci_sun4v_msi_getmsiq(pbm->devhandle, msi_num, &msiqid);
 999         if (err) {
1000                 printk(KERN_ERR "%s: getmsiq gives error %lu\n",
1001                        pbm->name, err);
1002                 return;
1003         }
1004
1005         pci_sun4v_msi_setvalid(pbm->devhandle, msi_num, HV_MSIVALID_INVALID);
1006         pci_sun4v_msiq_setvalid(pbm->devhandle, msiqid, HV_MSIQ_INVALID);
1007
1008         free_msi(pbm, msi_num);
1009
1010         /* The sun4v_destroy_msi() will liberate the devino and thus the MSIQ
1011          * allocation.
1012          */
1013         sun4v_destroy_msi(virt_irq);
1014 }
1015
1016 static void pci_sun4v_msi_init(struct pci_pbm_info *pbm)
1017 {
1018         const u32 *val;
1019         int len;
1020
1021         val = of_get_property(pbm->prom_node, "#msi-eqs", &len);
1022         if (!val || len != 4)
1023                 goto no_msi;
1024         pbm->msiq_num = *val;
1025         if (pbm->msiq_num) {
1026                 const struct msiq_prop {
1027                         u32 first_msiq;
1028                         u32 num_msiq;
1029                         u32 first_devino;
1030                 } *mqp;
1031                 const struct msi_range_prop {
1032                         u32 first_msi;
1033                         u32 num_msi;
1034                 } *mrng;
1035                 const struct addr_range_prop {
1036                         u32 msi32_high;
1037                         u32 msi32_low;
1038                         u32 msi32_len;
1039                         u32 msi64_high;
1040                         u32 msi64_low;
1041                         u32 msi64_len;
1042                 } *arng;
1043
1044                 val = of_get_property(pbm->prom_node, "msi-eq-size", &len);
1045                 if (!val || len != 4)
1046                         goto no_msi;
1047
1048                 pbm->msiq_ent_count = *val;
1049
1050                 mqp = of_get_property(pbm->prom_node,
1051                                       "msi-eq-to-devino", &len);
1052                 if (!mqp || len != sizeof(struct msiq_prop))
1053                         goto no_msi;
1054
1055                 pbm->msiq_first = mqp->first_msiq;
1056                 pbm->msiq_first_devino = mqp->first_devino;
1057
1058                 val = of_get_property(pbm->prom_node, "#msi", &len);
1059                 if (!val || len != 4)
1060                         goto no_msi;
1061                 pbm->msi_num = *val;
1062
1063                 mrng = of_get_property(pbm->prom_node, "msi-ranges", &len);
1064                 if (!mrng || len != sizeof(struct msi_range_prop))
1065                         goto no_msi;
1066                 pbm->msi_first = mrng->first_msi;
1067
1068                 val = of_get_property(pbm->prom_node, "msi-data-mask", &len);
1069                 if (!val || len != 4)
1070                         goto no_msi;
1071                 pbm->msi_data_mask = *val;
1072
1073                 val = of_get_property(pbm->prom_node, "msix-data-width", &len);
1074                 if (!val || len != 4)
1075                         goto no_msi;
1076                 pbm->msix_data_width = *val;
1077
1078                 arng = of_get_property(pbm->prom_node, "msi-address-ranges",
1079                                        &len);
1080                 if (!arng || len != sizeof(struct addr_range_prop))
1081                         goto no_msi;
1082                 pbm->msi32_start = ((u64)arng->msi32_high << 32) |
1083                         (u64) arng->msi32_low;
1084                 pbm->msi64_start = ((u64)arng->msi64_high << 32) |
1085                         (u64) arng->msi64_low;
1086                 pbm->msi32_len = arng->msi32_len;
1087                 pbm->msi64_len = arng->msi64_len;
1088
1089                 if (msi_bitmap_alloc(pbm))
1090                         goto no_msi;
1091
1092                 if (msi_queue_alloc(pbm)) {
1093                         msi_bitmap_free(pbm);
1094                         goto no_msi;
1095                 }
1096
1097                 printk(KERN_INFO "%s: MSI Queue first[%u] num[%u] count[%u] "
1098                        "devino[0x%x]\n",
1099                        pbm->name,
1100                        pbm->msiq_first, pbm->msiq_num,
1101                        pbm->msiq_ent_count,
1102                        pbm->msiq_first_devino);
1103                 printk(KERN_INFO "%s: MSI first[%u] num[%u] mask[0x%x] "
1104                        "width[%u]\n",
1105                        pbm->name,
1106                        pbm->msi_first, pbm->msi_num, pbm->msi_data_mask,
1107                        pbm->msix_data_width);
1108                 printk(KERN_INFO "%s: MSI addr32[0x%lx:0x%x] "
1109                        "addr64[0x%lx:0x%x]\n",
1110                        pbm->name,
1111                        pbm->msi32_start, pbm->msi32_len,
1112                        pbm->msi64_start, pbm->msi64_len);
1113                 printk(KERN_INFO "%s: MSI queues at RA [%p]\n",
1114                        pbm->name,
1115                        pbm->msi_queues);
1116         }
1117         pbm->setup_msi_irq = pci_sun4v_setup_msi_irq;
1118         pbm->teardown_msi_irq = pci_sun4v_teardown_msi_irq;
1119
1120         return;
1121
1122 no_msi:
1123         pbm->msiq_num = 0;
1124         printk(KERN_INFO "%s: No MSI support.\n", pbm->name);
1125 }
1126 #else /* CONFIG_PCI_MSI */
1127 static void pci_sun4v_msi_init(struct pci_pbm_info *pbm)
1128 {
1129 }
1130 #endif /* !(CONFIG_PCI_MSI) */
1131
1132 static void __init pci_sun4v_pbm_init(struct pci_controller_info *p, struct device_node *dp, u32 devhandle)
1133 {
1134         struct pci_pbm_info *pbm;
1135
1136         if (devhandle & 0x40)
1137                 pbm = &p->pbm_B;
1138         else
1139                 pbm = &p->pbm_A;
1140
1141         pbm->next = pci_pbm_root;
1142         pci_pbm_root = pbm;
1143
1144         pbm->scan_bus = pci_sun4v_scan_bus;
1145         pbm->pci_ops = &sun4v_pci_ops;
1146         pbm->config_space_reg_bits = 12;
1147
1148         pbm->index = pci_num_pbms++;
1149
1150         pbm->parent = p;
1151         pbm->prom_node = dp;
1152
1153         pbm->devhandle = devhandle;
1154
1155         pbm->name = dp->full_name;
1156
1157         printk("%s: SUN4V PCI Bus Module\n", pbm->name);
1158
1159         pci_determine_mem_io_space(pbm);
1160
1161         pci_get_pbm_props(pbm);
1162         pci_sun4v_iommu_init(pbm);
1163         pci_sun4v_msi_init(pbm);
1164 }
1165
1166 void __init sun4v_pci_init(struct device_node *dp, char *model_name)
1167 {
1168         static int hvapi_negotiated = 0;
1169         struct pci_controller_info *p;
1170         struct pci_pbm_info *pbm;
1171         struct iommu *iommu;
1172         struct property *prop;
1173         struct linux_prom64_registers *regs;
1174         u32 devhandle;
1175         int i;
1176
1177         if (!hvapi_negotiated++) {
1178                 int err = sun4v_hvapi_register(HV_GRP_PCI,
1179                                                vpci_major,
1180                                                &vpci_minor);
1181
1182                 if (err) {
1183                         prom_printf("SUN4V_PCI: Could not register hvapi, "
1184                                     "err=%d\n", err);
1185                         prom_halt();
1186                 }
1187                 printk("SUN4V_PCI: Registered hvapi major[%lu] minor[%lu]\n",
1188                        vpci_major, vpci_minor);
1189         }
1190
1191         prop = of_find_property(dp, "reg", NULL);
1192         regs = prop->value;
1193
1194         devhandle = (regs->phys_addr >> 32UL) & 0x0fffffff;
1195
1196         for (pbm = pci_pbm_root; pbm; pbm = pbm->next) {
1197                 if (pbm->devhandle == (devhandle ^ 0x40)) {
1198                         pci_sun4v_pbm_init(pbm->parent, dp, devhandle);
1199                         return;
1200                 }
1201         }
1202
1203         for_each_possible_cpu(i) {
1204                 unsigned long page = get_zeroed_page(GFP_ATOMIC);
1205
1206                 if (!page)
1207                         goto fatal_memory_error;
1208
1209                 per_cpu(pci_iommu_batch, i).pglist = (u64 *) page;
1210         }
1211
1212         p = kzalloc(sizeof(struct pci_controller_info), GFP_ATOMIC);
1213         if (!p)
1214                 goto fatal_memory_error;
1215
1216         iommu = kzalloc(sizeof(struct iommu), GFP_ATOMIC);
1217         if (!iommu)
1218                 goto fatal_memory_error;
1219
1220         p->pbm_A.iommu = iommu;
1221
1222         iommu = kzalloc(sizeof(struct iommu), GFP_ATOMIC);
1223         if (!iommu)
1224                 goto fatal_memory_error;
1225
1226         p->pbm_B.iommu = iommu;
1227
1228         /* Like PSYCHO and SCHIZO we have a 2GB aligned area
1229          * for memory space.
1230          */
1231         pci_memspace_mask = 0x7fffffffUL;
1232
1233         pci_sun4v_pbm_init(p, dp, devhandle);
1234         return;
1235
1236 fatal_memory_error:
1237         prom_printf("SUN4V_PCI: Fatal memory allocation error.\n");
1238         prom_halt();
1239 }