arch/x86/kernel/pci-dma.c

   1 #include <linux/dma-mapping.h>
   2 #include <linux/dma-debug.h>
   3 #include <linux/dmar.h>
   4 #include <linux/export.h>
   5 #include <linux/bootmem.h>
   6 #include <linux/gfp.h>
   7 #include <linux/pci.h>
   8 #include <linux/kmemleak.h>
   9
  10 #include <asm/proto.h>
  11 #include <asm/dma.h>
  12 #include <asm/iommu.h>
  13 #include <asm/gart.h>
  14 #include <asm/calgary.h>
  15 #include <asm/x86_init.h>
  16 #include <asm/iommu_table.h>
  17
  18 static int forbid_dac __read_mostly;
  19
  20 struct dma_map_ops *dma_ops = &nommu_dma_ops;
  21 EXPORT_SYMBOL(dma_ops);
  22
  23 static int iommu_sac_force __read_mostly;
  24
  25 #ifdef CONFIG_IOMMU_DEBUG
  26 int panic_on_overflow __read_mostly = 1;
  27 int force_iommu __read_mostly = 1;
  28 #else
  29 int panic_on_overflow __read_mostly = 0;
  30 int force_iommu __read_mostly = 0;
  31 #endif
  32
  33 int iommu_merge __read_mostly = 0;
  34
  35 int no_iommu __read_mostly;
  36 /* Set this to 1 if there is a HW IOMMU in the system */
  37 int iommu_detected __read_mostly = 0;
  38
  39 /*
  40  * This variable becomes 1 if iommu=pt is passed on the kernel command line.
  41  * If this variable is 1, IOMMU implementations do no DMA translation for
  42  * devices and allow every device to access to whole physical memory. This is
  43  * useful if a user wants to use an IOMMU only for KVM device assignment to
  44  * guests and not for driver dma translation.
  45  */
  46 int iommu_pass_through __read_mostly;
  47
  48 /*
  49  * Group multi-function PCI devices into a single device-group for the
  50  * iommu_device_group interface.  This tells the iommu driver to pretend
  51  * it cannot distinguish between functions of a device, exposing only one
  52  * group for the device.  Useful for disallowing use of individual PCI
  53  * functions from userspace drivers.
  54  */
  55 int iommu_group_mf __read_mostly;
  56
  57 extern struct iommu_table_entry __iommu_table[], __iommu_table_end[];
  58
  59 /* Dummy device used for NULL arguments (normally ISA). */
  60 struct device x86_dma_fallback_dev = {
  61         .init_name = "fallback device",
  62         .coherent_dma_mask = ISA_DMA_BIT_MASK,
  63         .dma_mask = &x86_dma_fallback_dev.coherent_dma_mask,
  64 };
  65 EXPORT_SYMBOL(x86_dma_fallback_dev);
  66
  67 /* Number of entries preallocated for DMA-API debugging */
  68 #define PREALLOC_DMA_DEBUG_ENTRIES       32768
  69
  70 int dma_set_mask(struct device *dev, u64 mask)
  71 {
  72         if (!dev->dma_mask || !dma_supported(dev, mask))
  73                 return -EIO;
  74
  75         *dev->dma_mask = mask;
  76
  77         return 0;
  78 }
  79 EXPORT_SYMBOL(dma_set_mask);
  80
  81 void __init pci_iommu_alloc(void)
  82 {
  83         struct iommu_table_entry *p;
  84
  85         sort_iommu_table(__iommu_table, __iommu_table_end);
  86         check_iommu_entries(__iommu_table, __iommu_table_end);
  87
  88         for (p = __iommu_table; p < __iommu_table_end; p++) {
  89                 if (p && p->detect && p->detect() > 0) {
  90                         p->flags |= IOMMU_DETECTED;
  91                         if (p->early_init)
  92                                 p->early_init();
  93                         if (p->flags & IOMMU_FINISH_IF_DETECTED)
  94                                 break;
  95                 }
  96         }
  97 }
  98 void *dma_generic_alloc_coherent(struct device *dev, size_t size,
  99                                  dma_addr_t *dma_addr, gfp_t flag,
 100                                  struct dma_attrs *attrs)
 101 {
 102         unsigned long dma_mask;
 103         struct page *page;
 104         dma_addr_t addr;
 105
 106         dma_mask = dma_alloc_coherent_mask(dev, flag);
 107
 108         flag |= __GFP_ZERO;
 109 again:
 110         page = alloc_pages_node(dev_to_node(dev), flag, get_order(size));
 111         if (!page)
 112                 return NULL;
 113
 114         addr = page_to_phys(page);
 115         if (addr + size > dma_mask) {
 116                 __free_pages(page, get_order(size));
 117
 118                 if (dma_mask < DMA_BIT_MASK(32) && !(flag & GFP_DMA)) {
 119                         flag = (flag & ~GFP_DMA32) | GFP_DMA;
 120                         goto again;
 121                 }
 122
 123                 return NULL;
 124         }
 125
 126         *dma_addr = addr;
 127         return page_address(page);
 128 }
 129
 130 /*
 131  * See <Documentation/x86/x86_64/boot-options.txt> for the iommu kernel
 132  * parameter documentation.
 133  */
 134 static __init int iommu_setup(char *p)
 135 {
 136         iommu_merge = 1;
 137
 138         if (!p)
 139                 return -EINVAL;
 140
 141         while (*p) {
 142                 if (!strncmp(p, "off", 3))
 143                         no_iommu = 1;
 144                 /* gart_parse_options has more force support */
 145                 if (!strncmp(p, "force", 5))
 146                         force_iommu = 1;
 147                 if (!strncmp(p, "noforce", 7)) {
 148                         iommu_merge = 0;
 149                         force_iommu = 0;
 150                 }
 151
 152                 if (!strncmp(p, "biomerge", 8)) {
 153                         iommu_merge = 1;
 154                         force_iommu = 1;
 155                 }
 156                 if (!strncmp(p, "panic", 5))
 157                         panic_on_overflow = 1;
 158                 if (!strncmp(p, "nopanic", 7))
 159                         panic_on_overflow = 0;
 160                 if (!strncmp(p, "merge", 5)) {
 161                         iommu_merge = 1;
 162                         force_iommu = 1;
 163                 }
 164                 if (!strncmp(p, "nomerge", 7))
 165                         iommu_merge = 0;
 166                 if (!strncmp(p, "forcesac", 8))
 167                         iommu_sac_force = 1;
 168                 if (!strncmp(p, "allowdac", 8))
 169                         forbid_dac = 0;
 170                 if (!strncmp(p, "nodac", 5))
 171                         forbid_dac = 1;
 172                 if (!strncmp(p, "usedac", 6)) {
 173                         forbid_dac = -1;
 174                         return 1;
 175                 }
 176 #ifdef CONFIG_SWIOTLB
 177                 if (!strncmp(p, "soft", 4))
 178                         swiotlb = 1;
 179 #endif
 180                 if (!strncmp(p, "pt", 2))
 181                         iommu_pass_through = 1;
 182                 if (!strncmp(p, "group_mf", 8))
 183                         iommu_group_mf = 1;
 184
 185                 gart_parse_options(p);
 186
 187 #ifdef CONFIG_CALGARY_IOMMU
 188                 if (!strncmp(p, "calgary", 7))
 189                         use_calgary = 1;
 190 #endif /* CONFIG_CALGARY_IOMMU */
 191
 192                 p += strcspn(p, ",");
 193                 if (*p == ',')
 194                         ++p;
 195         }
 196         return 0;
 197 }
 198 early_param("iommu", iommu_setup);
 199
 200 int dma_supported(struct device *dev, u64 mask)
 201 {
 202         struct dma_map_ops *ops = get_dma_ops(dev);
 203
 204 #ifdef CONFIG_PCI
 205         if (mask > 0xffffffff && forbid_dac > 0) {
 206                 dev_info(dev, "PCI: Disallowing DAC for device\n");
 207                 return 0;
 208         }
 209 #endif
 210
 211         if (ops->dma_supported)
 212                 return ops->dma_supported(dev, mask);
 213
 214         /* Copied from i386. Doesn't make much sense, because it will
 215            only work for pci_alloc_coherent.
 216            The caller just has to use GFP_DMA in this case. */
 217         if (mask < DMA_BIT_MASK(24))
 218                 return 0;
 219
 220         /* Tell the device to use SAC when IOMMU force is on.  This
 221            allows the driver to use cheaper accesses in some cases.
 222
 223            Problem with this is that if we overflow the IOMMU area and
 224            return DAC as fallback address the device may not handle it
 225            correctly.
 226
 227            As a special case some controllers have a 39bit address
 228            mode that is as efficient as 32bit (aic79xx). Don't force
 229            SAC for these.  Assume all masks <= 40 bits are of this
 230            type. Normally this doesn't make any difference, but gives
 231            more gentle handling of IOMMU overflow. */
 232         if (iommu_sac_force && (mask >= DMA_BIT_MASK(40))) {
 233                 dev_info(dev, "Force SAC with mask %Lx\n", mask);
 234                 return 0;
 235         }
 236
 237         return 1;
 238 }
 239 EXPORT_SYMBOL(dma_supported);
 240
 241 static int __init pci_iommu_init(void)
 242 {
 243         struct iommu_table_entry *p;
 244         dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
 245
 246 #ifdef CONFIG_PCI
 247         dma_debug_add_bus(&pci_bus_type);
 248 #endif
 249         x86_init.iommu.iommu_init();
 250
 251         for (p = __iommu_table; p < __iommu_table_end; p++) {
 252                 if (p && (p->flags & IOMMU_DETECTED) && p->late_init)
 253                         p->late_init();
 254         }
 255
 256         return 0;
 257 }
 258 /* Must execute after PCI subsystem */
 259 rootfs_initcall(pci_iommu_init);
 260
 261 #ifdef CONFIG_PCI
 262 /* Many VIA bridges seem to corrupt data for DAC. Disable it here */
 263
 264 static __devinit void via_no_dac(struct pci_dev *dev)
 265 {
 266         if (forbid_dac == 0) {
 267                 dev_info(&dev->dev, "disabling DAC on VIA PCI bridge\n");
 268                 forbid_dac = 1;
 269         }
 270 }
 271 DECLARE_PCI_FIXUP_CLASS_FINAL(PCI_VENDOR_ID_VIA, PCI_ANY_ID,
 272                                 PCI_CLASS_BRIDGE_PCI, 8, via_no_dac);
 273 #endif