arch/x86/kernel/pci-dma.c

   1 #include <linux/dma-mapping.h>
   2 #include <linux/dmar.h>
   3 #include <linux/bootmem.h>
   4 #include <linux/pci.h>
   5
   6 #include <asm/proto.h>
   7 #include <asm/dma.h>
   8 #include <asm/iommu.h>
   9 #include <asm/calgary.h>
  10 #include <asm/amd_iommu.h>
  11
  12 static int forbid_dac __read_mostly;
  13
  14 struct dma_mapping_ops *dma_ops;
  15 EXPORT_SYMBOL(dma_ops);
  16
  17 static int iommu_sac_force __read_mostly;
  18
  19 #ifdef CONFIG_IOMMU_DEBUG
  20 int panic_on_overflow __read_mostly = 1;
  21 int force_iommu __read_mostly = 1;
  22 #else
  23 int panic_on_overflow __read_mostly = 0;
  24 int force_iommu __read_mostly = 0;
  25 #endif
  26
  27 int iommu_merge __read_mostly = 0;
  28
  29 int no_iommu __read_mostly;
  30 /* Set this to 1 if there is a HW IOMMU in the system */
  31 int iommu_detected __read_mostly = 0;
  32
  33 /* This tells the BIO block layer to assume merging. Default to off
  34    because we cannot guarantee merging later. */
  35 int iommu_bio_merge __read_mostly = 0;
  36 EXPORT_SYMBOL(iommu_bio_merge);
  37
  38 dma_addr_t bad_dma_address __read_mostly = 0;
  39 EXPORT_SYMBOL(bad_dma_address);
  40
  41 /* Dummy device used for NULL arguments (normally ISA). Better would
  42    be probably a smaller DMA mask, but this is bug-to-bug compatible
  43    to older i386. */
  44 struct device fallback_dev = {
  45         .bus_id = "fallback device",
  46         .coherent_dma_mask = DMA_32BIT_MASK,
  47         .dma_mask = &fallback_dev.coherent_dma_mask,
  48 };
  49
  50 int dma_set_mask(struct device *dev, u64 mask)
  51 {
  52         if (!dev->dma_mask || !dma_supported(dev, mask))
  53                 return -EIO;
  54
  55         *dev->dma_mask = mask;
  56
  57         return 0;
  58 }
  59 EXPORT_SYMBOL(dma_set_mask);
  60
  61 #ifdef CONFIG_X86_64
  62 static __initdata void *dma32_bootmem_ptr;
  63 static unsigned long dma32_bootmem_size __initdata = (128ULL<<20);
  64
  65 static int __init parse_dma32_size_opt(char *p)
  66 {
  67         if (!p)
  68                 return -EINVAL;
  69         dma32_bootmem_size = memparse(p, &p);
  70         return 0;
  71 }
  72 early_param("dma32_size", parse_dma32_size_opt);
  73
  74 void __init dma32_reserve_bootmem(void)
  75 {
  76         unsigned long size, align;
  77         if (max_pfn <= MAX_DMA32_PFN)
  78                 return;
  79
  80         /*
  81          * check aperture_64.c allocate_aperture() for reason about
  82          * using 512M as goal
  83          */
  84         align = 64ULL<<20;
  85         size = round_up(dma32_bootmem_size, align);
  86         dma32_bootmem_ptr = __alloc_bootmem_nopanic(size, align,
  87                                  512ULL<<20);
  88         if (dma32_bootmem_ptr)
  89                 dma32_bootmem_size = size;
  90         else
  91                 dma32_bootmem_size = 0;
  92 }
  93 static void __init dma32_free_bootmem(void)
  94 {
  95
  96         if (max_pfn <= MAX_DMA32_PFN)
  97                 return;
  98
  99         if (!dma32_bootmem_ptr)
 100                 return;
 101
 102         free_bootmem(__pa(dma32_bootmem_ptr), dma32_bootmem_size);
 103
 104         dma32_bootmem_ptr = NULL;
 105         dma32_bootmem_size = 0;
 106 }
 107
 108 void __init pci_iommu_alloc(void)
 109 {
 110         /* free the range so iommu could get some range less than 4G */
 111         dma32_free_bootmem();
 112         /*
 113          * The order of these functions is important for
 114          * fall-back/fail-over reasons
 115          */
 116         gart_iommu_hole_init();
 117
 118         detect_calgary();
 119
 120         detect_intel_iommu();
 121
 122         amd_iommu_detect();
 123
 124         pci_swiotlb_init();
 125 }
 126
 127 unsigned long iommu_num_pages(unsigned long addr, unsigned long len)
 128 {
 129         unsigned long size = roundup((addr & ~PAGE_MASK) + len, PAGE_SIZE);
 130
 131         return size >> PAGE_SHIFT;
 132 }
 133 EXPORT_SYMBOL(iommu_num_pages);
 134 #endif
 135
 136 /*
 137  * See <Documentation/x86_64/boot-options.txt> for the iommu kernel parameter
 138  * documentation.
 139  */
 140 static __init int iommu_setup(char *p)
 141 {
 142         iommu_merge = 1;
 143
 144         if (!p)
 145                 return -EINVAL;
 146
 147         while (*p) {
 148                 if (!strncmp(p, "off", 3))
 149                         no_iommu = 1;
 150                 /* gart_parse_options has more force support */
 151                 if (!strncmp(p, "force", 5))
 152                         force_iommu = 1;
 153                 if (!strncmp(p, "noforce", 7)) {
 154                         iommu_merge = 0;
 155                         force_iommu = 0;
 156                 }
 157
 158                 if (!strncmp(p, "biomerge", 8)) {
 159                         iommu_bio_merge = 4096;
 160                         iommu_merge = 1;
 161                         force_iommu = 1;
 162                 }
 163                 if (!strncmp(p, "panic", 5))
 164                         panic_on_overflow = 1;
 165                 if (!strncmp(p, "nopanic", 7))
 166                         panic_on_overflow = 0;
 167                 if (!strncmp(p, "merge", 5)) {
 168                         iommu_merge = 1;
 169                         force_iommu = 1;
 170                 }
 171                 if (!strncmp(p, "nomerge", 7))
 172                         iommu_merge = 0;
 173                 if (!strncmp(p, "forcesac", 8))
 174                         iommu_sac_force = 1;
 175                 if (!strncmp(p, "allowdac", 8))
 176                         forbid_dac = 0;
 177                 if (!strncmp(p, "nodac", 5))
 178                         forbid_dac = -1;
 179                 if (!strncmp(p, "usedac", 6)) {
 180                         forbid_dac = -1;
 181                         return 1;
 182                 }
 183 #ifdef CONFIG_SWIOTLB
 184                 if (!strncmp(p, "soft", 4))
 185                         swiotlb = 1;
 186 #endif
 187
 188                 gart_parse_options(p);
 189
 190 #ifdef CONFIG_CALGARY_IOMMU
 191                 if (!strncmp(p, "calgary", 7))
 192                         use_calgary = 1;
 193 #endif /* CONFIG_CALGARY_IOMMU */
 194
 195                 p += strcspn(p, ",");
 196                 if (*p == ',')
 197                         ++p;
 198         }
 199         return 0;
 200 }
 201 early_param("iommu", iommu_setup);
 202
 203 int dma_supported(struct device *dev, u64 mask)
 204 {
 205         struct dma_mapping_ops *ops = get_dma_ops(dev);
 206
 207 #ifdef CONFIG_PCI
 208         if (mask > 0xffffffff && forbid_dac > 0) {
 209                 dev_info(dev, "PCI: Disallowing DAC for device\n");
 210                 return 0;
 211         }
 212 #endif
 213
 214         if (ops->dma_supported)
 215                 return ops->dma_supported(dev, mask);
 216
 217         /* Copied from i386. Doesn't make much sense, because it will
 218            only work for pci_alloc_coherent.
 219            The caller just has to use GFP_DMA in this case. */
 220         if (mask < DMA_24BIT_MASK)
 221                 return 0;
 222
 223         /* Tell the device to use SAC when IOMMU force is on.  This
 224            allows the driver to use cheaper accesses in some cases.
 225
 226            Problem with this is that if we overflow the IOMMU area and
 227            return DAC as fallback address the device may not handle it
 228            correctly.
 229
 230            As a special case some controllers have a 39bit address
 231            mode that is as efficient as 32bit (aic79xx). Don't force
 232            SAC for these.  Assume all masks <= 40 bits are of this
 233            type. Normally this doesn't make any difference, but gives
 234            more gentle handling of IOMMU overflow. */
 235         if (iommu_sac_force && (mask >= DMA_40BIT_MASK)) {
 236                 dev_info(dev, "Force SAC with mask %Lx\n", mask);
 237                 return 0;
 238         }
 239
 240         return 1;
 241 }
 242 EXPORT_SYMBOL(dma_supported);
 243
 244 /* Allocate DMA memory on node near device */
 245 static noinline struct page *
 246 dma_alloc_pages(struct device *dev, gfp_t gfp, unsigned order)
 247 {
 248         int node;
 249
 250         node = dev_to_node(dev);
 251
 252         return alloc_pages_node(node, gfp, order);
 253 }
 254
 255 /*
 256  * Allocate memory for a coherent mapping.
 257  */
 258 void *
 259 dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_handle,
 260                    gfp_t gfp)
 261 {
 262         struct dma_mapping_ops *ops = get_dma_ops(dev);
 263         void *memory = NULL;
 264         struct page *page;
 265         unsigned long dma_mask = 0;
 266         dma_addr_t bus;
 267         int noretry = 0;
 268
 269         /* ignore region specifiers */
 270         gfp &= ~(__GFP_DMA | __GFP_HIGHMEM | __GFP_DMA32);
 271
 272         if (dma_alloc_from_coherent(dev, size, dma_handle, &memory))
 273                 return memory;
 274
 275         if (!dev) {
 276                 dev = &fallback_dev;
 277                 gfp |= GFP_DMA;
 278         }
 279         dma_mask = dev->coherent_dma_mask;
 280         if (dma_mask == 0)
 281                 dma_mask = (gfp & GFP_DMA) ? DMA_24BIT_MASK : DMA_32BIT_MASK;
 282
 283         /* Device not DMA able */
 284         if (dev->dma_mask == NULL)
 285                 return NULL;
 286
 287         /* Don't invoke OOM killer or retry in lower 16MB DMA zone */
 288         if (gfp & __GFP_DMA)
 289                 noretry = 1;
 290
 291 #ifdef CONFIG_X86_64
 292         /* Why <=? Even when the mask is smaller than 4GB it is often
 293            larger than 16MB and in this case we have a chance of
 294            finding fitting memory in the next higher zone first. If
 295            not retry with true GFP_DMA. -AK */
 296         if (dma_mask <= DMA_32BIT_MASK && !(gfp & GFP_DMA)) {
 297                 gfp |= GFP_DMA32;
 298                 if (dma_mask < DMA_32BIT_MASK)
 299                         noretry = 1;
 300         }
 301 #endif
 302
 303  again:
 304         page = dma_alloc_pages(dev,
 305                 noretry ? gfp | __GFP_NORETRY : gfp, get_order(size));
 306         if (page == NULL)
 307                 return NULL;
 308
 309         {
 310                 int high, mmu;
 311                 bus = page_to_phys(page);
 312                 memory = page_address(page);
 313                 high = (bus + size) >= dma_mask;
 314                 mmu = high;
 315                 if (force_iommu && !(gfp & GFP_DMA))
 316                         mmu = 1;
 317                 else if (high) {
 318                         free_pages((unsigned long)memory,
 319                                    get_order(size));
 320
 321                         /* Don't use the 16MB ZONE_DMA unless absolutely
 322                            needed. It's better to use remapping first. */
 323                         if (dma_mask < DMA_32BIT_MASK && !(gfp & GFP_DMA)) {
 324                                 gfp = (gfp & ~GFP_DMA32) | GFP_DMA;
 325                                 goto again;
 326                         }
 327
 328                         /* Let low level make its own zone decisions */
 329                         gfp &= ~(GFP_DMA32|GFP_DMA);
 330
 331                         if (ops->alloc_coherent)
 332                                 return ops->alloc_coherent(dev, size,
 333                                                            dma_handle, gfp);
 334                         return NULL;
 335                 }
 336
 337                 memset(memory, 0, size);
 338                 if (!mmu) {
 339                         *dma_handle = bus;
 340                         return memory;
 341                 }
 342         }
 343
 344         if (ops->alloc_coherent) {
 345                 free_pages((unsigned long)memory, get_order(size));
 346                 gfp &= ~(GFP_DMA|GFP_DMA32);
 347                 return ops->alloc_coherent(dev, size, dma_handle, gfp);
 348         }
 349
 350         if (ops->map_simple) {
 351                 *dma_handle = ops->map_simple(dev, virt_to_phys(memory),
 352                                               size,
 353                                               PCI_DMA_BIDIRECTIONAL);
 354                 if (*dma_handle != bad_dma_address)
 355                         return memory;
 356         }
 357
 358         if (panic_on_overflow)
 359                 panic("dma_alloc_coherent: IOMMU overflow by %lu bytes\n",
 360                       (unsigned long)size);
 361         free_pages((unsigned long)memory, get_order(size));
 362         return NULL;
 363 }
 364 EXPORT_SYMBOL(dma_alloc_coherent);
 365
 366 /*
 367  * Unmap coherent memory.
 368  * The caller must ensure that the device has finished accessing the mapping.
 369  */
 370 void dma_free_coherent(struct device *dev, size_t size,
 371                          void *vaddr, dma_addr_t bus)
 372 {
 373         struct dma_mapping_ops *ops = get_dma_ops(dev);
 374
 375         int order = get_order(size);
 376         WARN_ON(irqs_disabled());       /* for portability */
 377         if (dma_release_from_coherent(dev, order, vaddr))
 378                 return;
 379         if (ops->unmap_single)
 380                 ops->unmap_single(dev, bus, size, 0);
 381         free_pages((unsigned long)vaddr, order);
 382 }
 383 EXPORT_SYMBOL(dma_free_coherent);
 384
 385 static int __init pci_iommu_init(void)
 386 {
 387         calgary_iommu_init();
 388
 389         intel_iommu_init();
 390
 391         amd_iommu_init();
 392
 393         gart_iommu_init();
 394
 395         no_iommu_init();
 396         return 0;
 397 }
 398
 399 void pci_iommu_shutdown(void)
 400 {
 401         gart_iommu_shutdown();
 402 }
 403 /* Must execute after PCI subsystem */
 404 fs_initcall(pci_iommu_init);
 405
 406 #ifdef CONFIG_PCI
 407 /* Many VIA bridges seem to corrupt data for DAC. Disable it here */
 408
 409 static __devinit void via_no_dac(struct pci_dev *dev)
 410 {
 411         if ((dev->class >> 8) == PCI_CLASS_BRIDGE_PCI && forbid_dac == 0) {
 412                 printk(KERN_INFO "PCI: VIA PCI bridge detected."
 413                                  "Disabling DAC.\n");
 414                 forbid_dac = 1;
 415         }
 416 }
 417 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_VIA, PCI_ANY_ID, via_no_dac);
 418 #endif