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 x86_dma_fallback_dev = {
  45         .bus_id = "fallback device",
  46         .coherent_dma_mask = DMA_32BIT_MASK,
  47         .dma_mask = &x86_dma_fallback_dev.coherent_dma_mask,
  48 };
  49 EXPORT_SYMBOL(x86_dma_fallback_dev);
  50
  51 int dma_set_mask(struct device *dev, u64 mask)
  52 {
  53         if (!dev->dma_mask || !dma_supported(dev, mask))
  54                 return -EIO;
  55
  56         *dev->dma_mask = mask;
  57
  58         return 0;
  59 }
  60 EXPORT_SYMBOL(dma_set_mask);
  61
  62 #ifdef CONFIG_X86_64
  63 static __initdata void *dma32_bootmem_ptr;
  64 static unsigned long dma32_bootmem_size __initdata = (128ULL<<20);
  65
  66 static int __init parse_dma32_size_opt(char *p)
  67 {
  68         if (!p)
  69                 return -EINVAL;
  70         dma32_bootmem_size = memparse(p, &p);
  71         return 0;
  72 }
  73 early_param("dma32_size", parse_dma32_size_opt);
  74
  75 void __init dma32_reserve_bootmem(void)
  76 {
  77         unsigned long size, align;
  78         if (max_pfn <= MAX_DMA32_PFN)
  79                 return;
  80
  81         /*
  82          * check aperture_64.c allocate_aperture() for reason about
  83          * using 512M as goal
  84          */
  85         align = 64ULL<<20;
  86         size = roundup(dma32_bootmem_size, align);
  87         dma32_bootmem_ptr = __alloc_bootmem_nopanic(size, align,
  88                                  512ULL<<20);
  89         if (dma32_bootmem_ptr)
  90                 dma32_bootmem_size = size;
  91         else
  92                 dma32_bootmem_size = 0;
  93 }
  94 static void __init dma32_free_bootmem(void)
  95 {
  96
  97         if (max_pfn <= MAX_DMA32_PFN)
  98                 return;
  99
 100         if (!dma32_bootmem_ptr)
 101                 return;
 102
 103         free_bootmem(__pa(dma32_bootmem_ptr), dma32_bootmem_size);
 104
 105         dma32_bootmem_ptr = NULL;
 106         dma32_bootmem_size = 0;
 107 }
 108
 109 void __init pci_iommu_alloc(void)
 110 {
 111         /* free the range so iommu could get some range less than 4G */
 112         dma32_free_bootmem();
 113         /*
 114          * The order of these functions is important for
 115          * fall-back/fail-over reasons
 116          */
 117         gart_iommu_hole_init();
 118
 119         detect_calgary();
 120
 121         detect_intel_iommu();
 122
 123         amd_iommu_detect();
 124
 125         pci_swiotlb_init();
 126 }
 127
 128 unsigned long iommu_num_pages(unsigned long addr, unsigned long len)
 129 {
 130         unsigned long size = roundup((addr & ~PAGE_MASK) + len, PAGE_SIZE);
 131
 132         return size >> PAGE_SHIFT;
 133 }
 134 EXPORT_SYMBOL(iommu_num_pages);
 135 #endif
 136
 137 void *dma_generic_alloc_coherent(struct device *dev, size_t size,
 138                                  dma_addr_t *dma_addr, gfp_t flag)
 139 {
 140         unsigned long dma_mask;
 141         struct page *page;
 142         dma_addr_t addr;
 143
 144         dma_mask = dma_alloc_coherent_mask(dev, flag);
 145
 146         flag |= __GFP_ZERO;
 147 again:
 148         page = alloc_pages_node(dev_to_node(dev), flag, get_order(size));
 149         if (!page)
 150                 return NULL;
 151
 152         addr = page_to_phys(page);
 153         if (!is_buffer_dma_capable(dma_mask, addr, size)) {
 154                 __free_pages(page, get_order(size));
 155
 156                 if (dma_mask < DMA_32BIT_MASK && !(flag & GFP_DMA)) {
 157                         flag = (flag & ~GFP_DMA32) | GFP_DMA;
 158                         goto again;
 159                 }
 160
 161                 return NULL;
 162         }
 163
 164         *dma_addr = addr;
 165         return page_address(page);
 166 }
 167
 168 /*
 169  * See <Documentation/x86_64/boot-options.txt> for the iommu kernel parameter
 170  * documentation.
 171  */
 172 static __init int iommu_setup(char *p)
 173 {
 174         iommu_merge = 1;
 175
 176         if (!p)
 177                 return -EINVAL;
 178
 179         while (*p) {
 180                 if (!strncmp(p, "off", 3))
 181                         no_iommu = 1;
 182                 /* gart_parse_options has more force support */
 183                 if (!strncmp(p, "force", 5))
 184                         force_iommu = 1;
 185                 if (!strncmp(p, "noforce", 7)) {
 186                         iommu_merge = 0;
 187                         force_iommu = 0;
 188                 }
 189
 190                 if (!strncmp(p, "biomerge", 8)) {
 191                         iommu_bio_merge = 4096;
 192                         iommu_merge = 1;
 193                         force_iommu = 1;
 194                 }
 195                 if (!strncmp(p, "panic", 5))
 196                         panic_on_overflow = 1;
 197                 if (!strncmp(p, "nopanic", 7))
 198                         panic_on_overflow = 0;
 199                 if (!strncmp(p, "merge", 5)) {
 200                         iommu_merge = 1;
 201                         force_iommu = 1;
 202                 }
 203                 if (!strncmp(p, "nomerge", 7))
 204                         iommu_merge = 0;
 205                 if (!strncmp(p, "forcesac", 8))
 206                         iommu_sac_force = 1;
 207                 if (!strncmp(p, "allowdac", 8))
 208                         forbid_dac = 0;
 209                 if (!strncmp(p, "nodac", 5))
 210                         forbid_dac = -1;
 211                 if (!strncmp(p, "usedac", 6)) {
 212                         forbid_dac = -1;
 213                         return 1;
 214                 }
 215 #ifdef CONFIG_SWIOTLB
 216                 if (!strncmp(p, "soft", 4))
 217                         swiotlb = 1;
 218 #endif
 219
 220                 gart_parse_options(p);
 221
 222 #ifdef CONFIG_CALGARY_IOMMU
 223                 if (!strncmp(p, "calgary", 7))
 224                         use_calgary = 1;
 225 #endif /* CONFIG_CALGARY_IOMMU */
 226
 227                 p += strcspn(p, ",");
 228                 if (*p == ',')
 229                         ++p;
 230         }
 231         return 0;
 232 }
 233 early_param("iommu", iommu_setup);
 234
 235 int dma_supported(struct device *dev, u64 mask)
 236 {
 237         struct dma_mapping_ops *ops = get_dma_ops(dev);
 238
 239 #ifdef CONFIG_PCI
 240         if (mask > 0xffffffff && forbid_dac > 0) {
 241                 dev_info(dev, "PCI: Disallowing DAC for device\n");
 242                 return 0;
 243         }
 244 #endif
 245
 246         if (ops->dma_supported)
 247                 return ops->dma_supported(dev, mask);
 248
 249         /* Copied from i386. Doesn't make much sense, because it will
 250            only work for pci_alloc_coherent.
 251            The caller just has to use GFP_DMA in this case. */
 252         if (mask < DMA_24BIT_MASK)
 253                 return 0;
 254
 255         /* Tell the device to use SAC when IOMMU force is on.  This
 256            allows the driver to use cheaper accesses in some cases.
 257
 258            Problem with this is that if we overflow the IOMMU area and
 259            return DAC as fallback address the device may not handle it
 260            correctly.
 261
 262            As a special case some controllers have a 39bit address
 263            mode that is as efficient as 32bit (aic79xx). Don't force
 264            SAC for these.  Assume all masks <= 40 bits are of this
 265            type. Normally this doesn't make any difference, but gives
 266            more gentle handling of IOMMU overflow. */
 267         if (iommu_sac_force && (mask >= DMA_40BIT_MASK)) {
 268                 dev_info(dev, "Force SAC with mask %Lx\n", mask);
 269                 return 0;
 270         }
 271
 272         return 1;
 273 }
 274 EXPORT_SYMBOL(dma_supported);
 275
 276 static int __init pci_iommu_init(void)
 277 {
 278         calgary_iommu_init();
 279
 280         intel_iommu_init();
 281
 282         amd_iommu_init();
 283
 284         gart_iommu_init();
 285
 286         no_iommu_init();
 287         return 0;
 288 }
 289
 290 void pci_iommu_shutdown(void)
 291 {
 292         gart_iommu_shutdown();
 293 }
 294 /* Must execute after PCI subsystem */
 295 fs_initcall(pci_iommu_init);
 296
 297 #ifdef CONFIG_PCI
 298 /* Many VIA bridges seem to corrupt data for DAC. Disable it here */
 299
 300 static __devinit void via_no_dac(struct pci_dev *dev)
 301 {
 302         if ((dev->class >> 8) == PCI_CLASS_BRIDGE_PCI && forbid_dac == 0) {
 303                 printk(KERN_INFO "PCI: VIA PCI bridge detected."
 304                                  "Disabling DAC.\n");
 305                 forbid_dac = 1;
 306         }
 307 }
 308 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_VIA, PCI_ANY_ID, via_no_dac);
 309 #endif