arch/x86/kernel/pci-gart_64.c

   1 /*
   2  * Dynamic DMA mapping support for AMD Hammer.
   3  *
   4  * Use the integrated AGP GART in the Hammer northbridge as an IOMMU for PCI.
   5  * This allows to use PCI devices that only support 32bit addresses on systems
   6  * with more than 4GB.
   7  *
   8  * See Documentation/DMA-mapping.txt for the interface specification.
   9  *
  10  * Copyright 2002 Andi Kleen, SuSE Labs.
  11  * Subject to the GNU General Public License v2 only.
  12  */
  13
  14 #include <linux/types.h>
  15 #include <linux/ctype.h>
  16 #include <linux/agp_backend.h>
  17 #include <linux/init.h>
  18 #include <linux/mm.h>
  19 #include <linux/string.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/pci.h>
  22 #include <linux/module.h>
  23 #include <linux/topology.h>
  24 #include <linux/interrupt.h>
  25 #include <linux/bitops.h>
  26 #include <linux/kdebug.h>
  27 #include <linux/scatterlist.h>
  28 #include <linux/iommu-helper.h>
  29 #include <asm/atomic.h>
  30 #include <asm/io.h>
  31 #include <asm/mtrr.h>
  32 #include <asm/pgtable.h>
  33 #include <asm/proto.h>
  34 #include <asm/gart.h>
  35 #include <asm/cacheflush.h>
  36 #include <asm/swiotlb.h>
  37 #include <asm/dma.h>
  38 #include <asm/k8.h>
  39
  40 static unsigned long iommu_bus_base;    /* GART remapping area (physical) */
  41 static unsigned long iommu_size;        /* size of remapping area bytes */
  42 static unsigned long iommu_pages;       /* .. and in pages */
  43
  44 static u32 *iommu_gatt_base;            /* Remapping table */
  45
  46 /*
  47  * If this is disabled the IOMMU will use an optimized flushing strategy
  48  * of only flushing when an mapping is reused. With it true the GART is
  49  * flushed for every mapping. Problem is that doing the lazy flush seems
  50  * to trigger bugs with some popular PCI cards, in particular 3ware (but
  51  * has been also also seen with Qlogic at least).
  52  */
  53 int iommu_fullflush = 1;
  54
  55 /* Allocation bitmap for the remapping area: */
  56 static DEFINE_SPINLOCK(iommu_bitmap_lock);
  57 /* Guarded by iommu_bitmap_lock: */
  58 static unsigned long *iommu_gart_bitmap;
  59
  60 static u32 gart_unmapped_entry;
  61
  62 #define GPTE_VALID    1
  63 #define GPTE_COHERENT 2
  64 #define GPTE_ENCODE(x) \
  65         (((x) & 0xfffff000) | (((x) >> 32) << 4) | GPTE_VALID | GPTE_COHERENT)
  66 #define GPTE_DECODE(x) (((x) & 0xfffff000) | (((u64)(x) & 0xff0) << 28))
  67
  68 #define to_pages(addr, size) \
  69         (round_up(((addr) & ~PAGE_MASK) + (size), PAGE_SIZE) >> PAGE_SHIFT)
  70
  71 #define EMERGENCY_PAGES 32 /* = 128KB */
  72
  73 #ifdef CONFIG_AGP
  74 #define AGPEXTERN extern
  75 #else
  76 #define AGPEXTERN
  77 #endif
  78
  79 /* backdoor interface to AGP driver */
  80 AGPEXTERN int agp_memory_reserved;
  81 AGPEXTERN __u32 *agp_gatt_table;
  82
  83 static unsigned long next_bit;  /* protected by iommu_bitmap_lock */
  84 static int need_flush;          /* global flush state. set for each gart wrap */
  85
  86 static unsigned long alloc_iommu(struct device *dev, int size)
  87 {
  88         unsigned long offset, flags;
  89         unsigned long boundary_size;
  90         unsigned long base_index;
  91
  92         base_index = ALIGN(iommu_bus_base & dma_get_seg_boundary(dev),
  93                            PAGE_SIZE) >> PAGE_SHIFT;
  94         boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
  95                               PAGE_SIZE) >> PAGE_SHIFT;
  96
  97         spin_lock_irqsave(&iommu_bitmap_lock, flags);
  98         offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, next_bit,
  99                                   size, base_index, boundary_size, 0);
 100         if (offset == -1) {
 101                 need_flush = 1;
 102                 offset = iommu_area_alloc(iommu_gart_bitmap, iommu_pages, 0,
 103                                           size, base_index, boundary_size, 0);
 104         }
 105         if (offset != -1) {
 106                 set_bit_string(iommu_gart_bitmap, offset, size);
 107                 next_bit = offset+size;
 108                 if (next_bit >= iommu_pages) {
 109                         next_bit = 0;
 110                         need_flush = 1;
 111                 }
 112         }
 113         if (iommu_fullflush)
 114                 need_flush = 1;
 115         spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
 116
 117         return offset;
 118 }
 119
 120 static void free_iommu(unsigned long offset, int size)
 121 {
 122         unsigned long flags;
 123
 124         spin_lock_irqsave(&iommu_bitmap_lock, flags);
 125         iommu_area_free(iommu_gart_bitmap, offset, size);
 126         spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
 127 }
 128
 129 /*
 130  * Use global flush state to avoid races with multiple flushers.
 131  */
 132 static void flush_gart(void)
 133 {
 134         unsigned long flags;
 135
 136         spin_lock_irqsave(&iommu_bitmap_lock, flags);
 137         if (need_flush) {
 138                 k8_flush_garts();
 139                 need_flush = 0;
 140         }
 141         spin_unlock_irqrestore(&iommu_bitmap_lock, flags);
 142 }
 143
 144 #ifdef CONFIG_IOMMU_LEAK
 145
 146 #define SET_LEAK(x)                                                     \
 147         do {                                                            \
 148                 if (iommu_leak_tab)                                     \
 149                         iommu_leak_tab[x] = __builtin_return_address(0);\
 150         } while (0)
 151
 152 #define CLEAR_LEAK(x)                                                   \
 153         do {                                                            \
 154                 if (iommu_leak_tab)                                     \
 155                         iommu_leak_tab[x] = NULL;                       \
 156         } while (0)
 157
 158 /* Debugging aid for drivers that don't free their IOMMU tables */
 159 static void **iommu_leak_tab;
 160 static int leak_trace;
 161 static int iommu_leak_pages = 20;
 162
 163 static void dump_leak(void)
 164 {
 165         int i;
 166         static int dump;
 167
 168         if (dump || !iommu_leak_tab)
 169                 return;
 170         dump = 1;
 171         show_stack(NULL, NULL);
 172
 173         /* Very crude. dump some from the end of the table too */
 174         printk(KERN_DEBUG "Dumping %d pages from end of IOMMU:\n",
 175                iommu_leak_pages);
 176         for (i = 0; i < iommu_leak_pages; i += 2) {
 177                 printk(KERN_DEBUG "%lu: ", iommu_pages-i);
 178                 printk_address((unsigned long) iommu_leak_tab[iommu_pages-i], 0);
 179                 printk(KERN_CONT "%c", (i+1)%2 == 0 ? '\n' : ' ');
 180         }
 181         printk(KERN_DEBUG "\n");
 182 }
 183 #else
 184 # define SET_LEAK(x)
 185 # define CLEAR_LEAK(x)
 186 #endif
 187
 188 static void iommu_full(struct device *dev, size_t size, int dir)
 189 {
 190         /*
 191          * Ran out of IOMMU space for this operation. This is very bad.
 192          * Unfortunately the drivers cannot handle this operation properly.
 193          * Return some non mapped prereserved space in the aperture and
 194          * let the Northbridge deal with it. This will result in garbage
 195          * in the IO operation. When the size exceeds the prereserved space
 196          * memory corruption will occur or random memory will be DMAed
 197          * out. Hopefully no network devices use single mappings that big.
 198          */
 199
 200         printk(KERN_ERR
 201                 "PCI-DMA: Out of IOMMU space for %lu bytes at device %s\n",
 202                 size, dev->bus_id);
 203
 204         if (size > PAGE_SIZE*EMERGENCY_PAGES) {
 205                 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
 206                         panic("PCI-DMA: Memory would be corrupted\n");
 207                 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
 208                         panic(KERN_ERR
 209                                 "PCI-DMA: Random memory would be DMAed\n");
 210         }
 211 #ifdef CONFIG_IOMMU_LEAK
 212         dump_leak();
 213 #endif
 214 }
 215
 216 static inline int
 217 need_iommu(struct device *dev, unsigned long addr, size_t size)
 218 {
 219         u64 mask = *dev->dma_mask;
 220         int high = addr + size > mask;
 221         int mmu = high;
 222
 223         if (force_iommu)
 224                 mmu = 1;
 225
 226         return mmu;
 227 }
 228
 229 static inline int
 230 nonforced_iommu(struct device *dev, unsigned long addr, size_t size)
 231 {
 232         u64 mask = *dev->dma_mask;
 233         int high = addr + size > mask;
 234         int mmu = high;
 235
 236         return mmu;
 237 }
 238
 239 /* Map a single continuous physical area into the IOMMU.
 240  * Caller needs to check if the iommu is needed and flush.
 241  */
 242 static dma_addr_t dma_map_area(struct device *dev, dma_addr_t phys_mem,
 243                                 size_t size, int dir)
 244 {
 245         unsigned long npages = to_pages(phys_mem, size);
 246         unsigned long iommu_page = alloc_iommu(dev, npages);
 247         int i;
 248
 249         if (iommu_page == -1) {
 250                 if (!nonforced_iommu(dev, phys_mem, size))
 251                         return phys_mem;
 252                 if (panic_on_overflow)
 253                         panic("dma_map_area overflow %lu bytes\n", size);
 254                 iommu_full(dev, size, dir);
 255                 return bad_dma_address;
 256         }
 257
 258         for (i = 0; i < npages; i++) {
 259                 iommu_gatt_base[iommu_page + i] = GPTE_ENCODE(phys_mem);
 260                 SET_LEAK(iommu_page + i);
 261                 phys_mem += PAGE_SIZE;
 262         }
 263         return iommu_bus_base + iommu_page*PAGE_SIZE + (phys_mem & ~PAGE_MASK);
 264 }
 265
 266 static dma_addr_t
 267 gart_map_simple(struct device *dev, phys_addr_t paddr, size_t size, int dir)
 268 {
 269         dma_addr_t map = dma_map_area(dev, paddr, size, dir);
 270
 271         flush_gart();
 272
 273         return map;
 274 }
 275
 276 /* Map a single area into the IOMMU */
 277 static dma_addr_t
 278 gart_map_single(struct device *dev, phys_addr_t paddr, size_t size, int dir)
 279 {
 280         unsigned long bus;
 281
 282         if (!dev)
 283                 dev = &fallback_dev;
 284
 285         if (!need_iommu(dev, paddr, size))
 286                 return paddr;
 287
 288         bus = gart_map_simple(dev, paddr, size, dir);
 289
 290         return bus;
 291 }
 292
 293 /*
 294  * Free a DMA mapping.
 295  */
 296 static void gart_unmap_single(struct device *dev, dma_addr_t dma_addr,
 297                               size_t size, int direction)
 298 {
 299         unsigned long iommu_page;
 300         int npages;
 301         int i;
 302
 303         if (dma_addr < iommu_bus_base + EMERGENCY_PAGES*PAGE_SIZE ||
 304             dma_addr >= iommu_bus_base + iommu_size)
 305                 return;
 306
 307         iommu_page = (dma_addr - iommu_bus_base)>>PAGE_SHIFT;
 308         npages = to_pages(dma_addr, size);
 309         for (i = 0; i < npages; i++) {
 310                 iommu_gatt_base[iommu_page + i] = gart_unmapped_entry;
 311                 CLEAR_LEAK(iommu_page + i);
 312         }
 313         free_iommu(iommu_page, npages);
 314 }
 315
 316 /*
 317  * Wrapper for pci_unmap_single working with scatterlists.
 318  */
 319 static void
 320 gart_unmap_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
 321 {
 322         struct scatterlist *s;
 323         int i;
 324
 325         for_each_sg(sg, s, nents, i) {
 326                 if (!s->dma_length || !s->length)
 327                         break;
 328                 gart_unmap_single(dev, s->dma_address, s->dma_length, dir);
 329         }
 330 }
 331
 332 /* Fallback for dma_map_sg in case of overflow */
 333 static int dma_map_sg_nonforce(struct device *dev, struct scatterlist *sg,
 334                                int nents, int dir)
 335 {
 336         struct scatterlist *s;
 337         int i;
 338
 339 #ifdef CONFIG_IOMMU_DEBUG
 340         printk(KERN_DEBUG "dma_map_sg overflow\n");
 341 #endif
 342
 343         for_each_sg(sg, s, nents, i) {
 344                 unsigned long addr = sg_phys(s);
 345
 346                 if (nonforced_iommu(dev, addr, s->length)) {
 347                         addr = dma_map_area(dev, addr, s->length, dir);
 348                         if (addr == bad_dma_address) {
 349                                 if (i > 0)
 350                                         gart_unmap_sg(dev, sg, i, dir);
 351                                 nents = 0;
 352                                 sg[0].dma_length = 0;
 353                                 break;
 354                         }
 355                 }
 356                 s->dma_address = addr;
 357                 s->dma_length = s->length;
 358         }
 359         flush_gart();
 360
 361         return nents;
 362 }
 363
 364 /* Map multiple scatterlist entries continuous into the first. */
 365 static int __dma_map_cont(struct device *dev, struct scatterlist *start,
 366                           int nelems, struct scatterlist *sout,
 367                           unsigned long pages)
 368 {
 369         unsigned long iommu_start = alloc_iommu(dev, pages);
 370         unsigned long iommu_page = iommu_start;
 371         struct scatterlist *s;
 372         int i;
 373
 374         if (iommu_start == -1)
 375                 return -1;
 376
 377         for_each_sg(start, s, nelems, i) {
 378                 unsigned long pages, addr;
 379                 unsigned long phys_addr = s->dma_address;
 380
 381                 BUG_ON(s != start && s->offset);
 382                 if (s == start) {
 383                         sout->dma_address = iommu_bus_base;
 384                         sout->dma_address += iommu_page*PAGE_SIZE + s->offset;
 385                         sout->dma_length = s->length;
 386                 } else {
 387                         sout->dma_length += s->length;
 388                 }
 389
 390                 addr = phys_addr;
 391                 pages = to_pages(s->offset, s->length);
 392                 while (pages--) {
 393                         iommu_gatt_base[iommu_page] = GPTE_ENCODE(addr);
 394                         SET_LEAK(iommu_page);
 395                         addr += PAGE_SIZE;
 396                         iommu_page++;
 397                 }
 398         }
 399         BUG_ON(iommu_page - iommu_start != pages);
 400
 401         return 0;
 402 }
 403
 404 static inline int
 405 dma_map_cont(struct device *dev, struct scatterlist *start, int nelems,
 406              struct scatterlist *sout, unsigned long pages, int need)
 407 {
 408         if (!need) {
 409                 BUG_ON(nelems != 1);
 410                 sout->dma_address = start->dma_address;
 411                 sout->dma_length = start->length;
 412                 return 0;
 413         }
 414         return __dma_map_cont(dev, start, nelems, sout, pages);
 415 }
 416
 417 /*
 418  * DMA map all entries in a scatterlist.
 419  * Merge chunks that have page aligned sizes into a continuous mapping.
 420  */
 421 static int
 422 gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
 423 {
 424         struct scatterlist *s, *ps, *start_sg, *sgmap;
 425         int need = 0, nextneed, i, out, start;
 426         unsigned long pages = 0;
 427         unsigned int seg_size;
 428         unsigned int max_seg_size;
 429
 430         if (nents == 0)
 431                 return 0;
 432
 433         if (!dev)
 434                 dev = &fallback_dev;
 435
 436         out = 0;
 437         start = 0;
 438         start_sg = sgmap = sg;
 439         seg_size = 0;
 440         max_seg_size = dma_get_max_seg_size(dev);
 441         ps = NULL; /* shut up gcc */
 442         for_each_sg(sg, s, nents, i) {
 443                 dma_addr_t addr = sg_phys(s);
 444
 445                 s->dma_address = addr;
 446                 BUG_ON(s->length == 0);
 447
 448                 nextneed = need_iommu(dev, addr, s->length);
 449
 450                 /* Handle the previous not yet processed entries */
 451                 if (i > start) {
 452                         /*
 453                          * Can only merge when the last chunk ends on a
 454                          * page boundary and the new one doesn't have an
 455                          * offset.
 456                          */
 457                         if (!iommu_merge || !nextneed || !need || s->offset ||
 458                             (s->length + seg_size > max_seg_size) ||
 459                             (ps->offset + ps->length) % PAGE_SIZE) {
 460                                 if (dma_map_cont(dev, start_sg, i - start,
 461                                                  sgmap, pages, need) < 0)
 462                                         goto error;
 463                                 out++;
 464                                 seg_size = 0;
 465                                 sgmap = sg_next(sgmap);
 466                                 pages = 0;
 467                                 start = i;
 468                                 start_sg = s;
 469                         }
 470                 }
 471
 472                 seg_size += s->length;
 473                 need = nextneed;
 474                 pages += to_pages(s->offset, s->length);
 475                 ps = s;
 476         }
 477         if (dma_map_cont(dev, start_sg, i - start, sgmap, pages, need) < 0)
 478                 goto error;
 479         out++;
 480         flush_gart();
 481         if (out < nents) {
 482                 sgmap = sg_next(sgmap);
 483                 sgmap->dma_length = 0;
 484         }
 485         return out;
 486
 487 error:
 488         flush_gart();
 489         gart_unmap_sg(dev, sg, out, dir);
 490
 491         /* When it was forced or merged try again in a dumb way */
 492         if (force_iommu || iommu_merge) {
 493                 out = dma_map_sg_nonforce(dev, sg, nents, dir);
 494                 if (out > 0)
 495                         return out;
 496         }
 497         if (panic_on_overflow)
 498                 panic("dma_map_sg: overflow on %lu pages\n", pages);
 499
 500         iommu_full(dev, pages << PAGE_SHIFT, dir);
 501         for_each_sg(sg, s, nents, i)
 502                 s->dma_address = bad_dma_address;
 503         return 0;
 504 }
 505
 506 static int no_agp;
 507
 508 static __init unsigned long check_iommu_size(unsigned long aper, u64 aper_size)
 509 {
 510         unsigned long a;
 511
 512         if (!iommu_size) {
 513                 iommu_size = aper_size;
 514                 if (!no_agp)
 515                         iommu_size /= 2;
 516         }
 517
 518         a = aper + iommu_size;
 519         iommu_size -= round_up(a, PMD_PAGE_SIZE) - a;
 520
 521         if (iommu_size < 64*1024*1024) {
 522                 printk(KERN_WARNING
 523                         "PCI-DMA: Warning: Small IOMMU %luMB."
 524                         " Consider increasing the AGP aperture in BIOS\n",
 525                                 iommu_size >> 20);
 526         }
 527
 528         return iommu_size;
 529 }
 530
 531 static __init unsigned read_aperture(struct pci_dev *dev, u32 *size)
 532 {
 533         unsigned aper_size = 0, aper_base_32, aper_order;
 534         u64 aper_base;
 535
 536         pci_read_config_dword(dev, 0x94, &aper_base_32);
 537         pci_read_config_dword(dev, 0x90, &aper_order);
 538         aper_order = (aper_order >> 1) & 7;
 539
 540         aper_base = aper_base_32 & 0x7fff;
 541         aper_base <<= 25;
 542
 543         aper_size = (32 * 1024 * 1024) << aper_order;
 544         if (aper_base + aper_size > 0x100000000UL || !aper_size)
 545                 aper_base = 0;
 546
 547         *size = aper_size;
 548         return aper_base;
 549 }
 550
 551 /*
 552  * Private Northbridge GATT initialization in case we cannot use the
 553  * AGP driver for some reason.
 554  */
 555 static __init int init_k8_gatt(struct agp_kern_info *info)
 556 {
 557         unsigned aper_size, gatt_size, new_aper_size;
 558         unsigned aper_base, new_aper_base;
 559         struct pci_dev *dev;
 560         void *gatt;
 561         int i;
 562
 563         printk(KERN_INFO "PCI-DMA: Disabling AGP.\n");
 564         aper_size = aper_base = info->aper_size = 0;
 565         dev = NULL;
 566         for (i = 0; i < num_k8_northbridges; i++) {
 567                 dev = k8_northbridges[i];
 568                 new_aper_base = read_aperture(dev, &new_aper_size);
 569                 if (!new_aper_base)
 570                         goto nommu;
 571
 572                 if (!aper_base) {
 573                         aper_size = new_aper_size;
 574                         aper_base = new_aper_base;
 575                 }
 576                 if (aper_size != new_aper_size || aper_base != new_aper_base)
 577                         goto nommu;
 578         }
 579         if (!aper_base)
 580                 goto nommu;
 581         info->aper_base = aper_base;
 582         info->aper_size = aper_size >> 20;
 583
 584         gatt_size = (aper_size >> PAGE_SHIFT) * sizeof(u32);
 585         gatt = (void *)__get_free_pages(GFP_KERNEL, get_order(gatt_size));
 586         if (!gatt)
 587                 panic("Cannot allocate GATT table");
 588         if (set_memory_uc((unsigned long)gatt, gatt_size >> PAGE_SHIFT))
 589                 panic("Could not set GART PTEs to uncacheable pages");
 590
 591         memset(gatt, 0, gatt_size);
 592         agp_gatt_table = gatt;
 593
 594         for (i = 0; i < num_k8_northbridges; i++) {
 595                 u32 gatt_reg;
 596                 u32 ctl;
 597
 598                 dev = k8_northbridges[i];
 599                 gatt_reg = __pa(gatt) >> 12;
 600                 gatt_reg <<= 4;
 601                 pci_write_config_dword(dev, 0x98, gatt_reg);
 602                 pci_read_config_dword(dev, 0x90, &ctl);
 603
 604                 ctl |= 1;
 605                 ctl &= ~((1<<4) | (1<<5));
 606
 607                 pci_write_config_dword(dev, 0x90, ctl);
 608         }
 609         flush_gart();
 610
 611         printk(KERN_INFO "PCI-DMA: aperture base @ %x size %u KB\n",
 612                aper_base, aper_size>>10);
 613         return 0;
 614
 615  nommu:
 616         /* Should not happen anymore */
 617         printk(KERN_WARNING "PCI-DMA: More than 4GB of RAM and no IOMMU\n"
 618                KERN_WARNING "falling back to iommu=soft.\n");
 619         return -1;
 620 }
 621
 622 extern int agp_amd64_init(void);
 623
 624 static const struct dma_mapping_ops gart_dma_ops = {
 625         .mapping_error                  = NULL,
 626         .map_single                     = gart_map_single,
 627         .map_simple                     = gart_map_simple,
 628         .unmap_single                   = gart_unmap_single,
 629         .sync_single_for_cpu            = NULL,
 630         .sync_single_for_device         = NULL,
 631         .sync_single_range_for_cpu      = NULL,
 632         .sync_single_range_for_device   = NULL,
 633         .sync_sg_for_cpu                = NULL,
 634         .sync_sg_for_device             = NULL,
 635         .map_sg                         = gart_map_sg,
 636         .unmap_sg                       = gart_unmap_sg,
 637 };
 638
 639 void gart_iommu_shutdown(void)
 640 {
 641         struct pci_dev *dev;
 642         int i;
 643
 644         if (no_agp && (dma_ops != &gart_dma_ops))
 645                 return;
 646
 647         for (i = 0; i < num_k8_northbridges; i++) {
 648                 u32 ctl;
 649
 650                 dev = k8_northbridges[i];
 651                 pci_read_config_dword(dev, 0x90, &ctl);
 652
 653                 ctl &= ~1;
 654
 655                 pci_write_config_dword(dev, 0x90, ctl);
 656         }
 657 }
 658
 659 void __init gart_iommu_init(void)
 660 {
 661         struct agp_kern_info info;
 662         unsigned long iommu_start;
 663         unsigned long aper_size;
 664         unsigned long scratch;
 665         long i;
 666
 667         if (cache_k8_northbridges() < 0 || num_k8_northbridges == 0) {
 668                 printk(KERN_INFO "PCI-GART: No AMD northbridge found.\n");
 669                 return;
 670         }
 671
 672 #ifndef CONFIG_AGP_AMD64
 673         no_agp = 1;
 674 #else
 675         /* Makefile puts PCI initialization via subsys_initcall first. */
 676         /* Add other K8 AGP bridge drivers here */
 677         no_agp = no_agp ||
 678                 (agp_amd64_init() < 0) ||
 679                 (agp_copy_info(agp_bridge, &info) < 0);
 680 #endif
 681
 682         if (swiotlb)
 683                 return;
 684
 685         /* Did we detect a different HW IOMMU? */
 686         if (iommu_detected && !gart_iommu_aperture)
 687                 return;
 688
 689         if (no_iommu ||
 690             (!force_iommu && end_pfn <= MAX_DMA32_PFN) ||
 691             !gart_iommu_aperture ||
 692             (no_agp && init_k8_gatt(&info) < 0)) {
 693                 if (end_pfn > MAX_DMA32_PFN) {
 694                         printk(KERN_WARNING "More than 4GB of memory "
 695                                           "but GART IOMMU not available.\n"
 696                                KERN_WARNING "falling back to iommu=soft.\n");
 697                 }
 698                 return;
 699         }
 700
 701         printk(KERN_INFO "PCI-DMA: using GART IOMMU.\n");
 702         aper_size = info.aper_size * 1024 * 1024;
 703         iommu_size = check_iommu_size(info.aper_base, aper_size);
 704         iommu_pages = iommu_size >> PAGE_SHIFT;
 705
 706         iommu_gart_bitmap = (void *) __get_free_pages(GFP_KERNEL,
 707                                                       get_order(iommu_pages/8));
 708         if (!iommu_gart_bitmap)
 709                 panic("Cannot allocate iommu bitmap\n");
 710         memset(iommu_gart_bitmap, 0, iommu_pages/8);
 711
 712 #ifdef CONFIG_IOMMU_LEAK
 713         if (leak_trace) {
 714                 iommu_leak_tab = (void *)__get_free_pages(GFP_KERNEL,
 715                                   get_order(iommu_pages*sizeof(void *)));
 716                 if (iommu_leak_tab)
 717                         memset(iommu_leak_tab, 0, iommu_pages * 8);
 718                 else
 719                         printk(KERN_DEBUG
 720                                "PCI-DMA: Cannot allocate leak trace area\n");
 721         }
 722 #endif
 723
 724         /*
 725          * Out of IOMMU space handling.
 726          * Reserve some invalid pages at the beginning of the GART.
 727          */
 728         set_bit_string(iommu_gart_bitmap, 0, EMERGENCY_PAGES);
 729
 730         agp_memory_reserved = iommu_size;
 731         printk(KERN_INFO
 732                "PCI-DMA: Reserving %luMB of IOMMU area in the AGP aperture\n",
 733                iommu_size >> 20);
 734
 735         iommu_start = aper_size - iommu_size;
 736         iommu_bus_base = info.aper_base + iommu_start;
 737         bad_dma_address = iommu_bus_base;
 738         iommu_gatt_base = agp_gatt_table + (iommu_start>>PAGE_SHIFT);
 739
 740         /*
 741          * Unmap the IOMMU part of the GART. The alias of the page is
 742          * always mapped with cache enabled and there is no full cache
 743          * coherency across the GART remapping. The unmapping avoids
 744          * automatic prefetches from the CPU allocating cache lines in
 745          * there. All CPU accesses are done via the direct mapping to
 746          * the backing memory. The GART address is only used by PCI
 747          * devices.
 748          */
 749         set_memory_np((unsigned long)__va(iommu_bus_base),
 750                                 iommu_size >> PAGE_SHIFT);
 751         /*
 752          * Tricky. The GART table remaps the physical memory range,
 753          * so the CPU wont notice potential aliases and if the memory
 754          * is remapped to UC later on, we might surprise the PCI devices
 755          * with a stray writeout of a cacheline. So play it sure and
 756          * do an explicit, full-scale wbinvd() _after_ having marked all
 757          * the pages as Not-Present:
 758          */
 759         wbinvd();
 760
 761         /*
 762          * Try to workaround a bug (thanks to BenH)
 763          * Set unmapped entries to a scratch page instead of 0.
 764          * Any prefetches that hit unmapped entries won't get an bus abort
 765          * then.
 766          */
 767         scratch = get_zeroed_page(GFP_KERNEL);
 768         if (!scratch)
 769                 panic("Cannot allocate iommu scratch page");
 770         gart_unmapped_entry = GPTE_ENCODE(__pa(scratch));
 771         for (i = EMERGENCY_PAGES; i < iommu_pages; i++)
 772                 iommu_gatt_base[i] = gart_unmapped_entry;
 773
 774         flush_gart();
 775         dma_ops = &gart_dma_ops;
 776 }
 777
 778 void __init gart_parse_options(char *p)
 779 {
 780         int arg;
 781
 782 #ifdef CONFIG_IOMMU_LEAK
 783         if (!strncmp(p, "leak", 4)) {
 784                 leak_trace = 1;
 785                 p += 4;
 786                 if (*p == '=') ++p;
 787                 if (isdigit(*p) && get_option(&p, &arg))
 788                         iommu_leak_pages = arg;
 789         }
 790 #endif
 791         if (isdigit(*p) && get_option(&p, &arg))
 792                 iommu_size = arg;
 793         if (!strncmp(p, "fullflush", 8))
 794                 iommu_fullflush = 1;
 795         if (!strncmp(p, "nofullflush", 11))
 796                 iommu_fullflush = 0;
 797         if (!strncmp(p, "noagp", 5))
 798                 no_agp = 1;
 799         if (!strncmp(p, "noaperture", 10))
 800                 fix_aperture = 0;
 801         /* duplicated from pci-dma.c */
 802         if (!strncmp(p, "force", 5))
 803                 gart_iommu_aperture_allowed = 1;
 804         if (!strncmp(p, "allowed", 7))
 805                 gart_iommu_aperture_allowed = 1;
 806         if (!strncmp(p, "memaper", 7)) {
 807                 fallback_aper_force = 1;
 808                 p += 7;
 809                 if (*p == '=') {
 810                         ++p;
 811                         if (get_option(&p, &arg))
 812                                 fallback_aper_order = arg;
 813                 }
 814         }
 815 }