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 extern struct iommu_table_entry __iommu_table[], __iommu_table_end[];
  49
  50 /* Dummy device used for NULL arguments (normally ISA). */
  51 struct device x86_dma_fallback_dev = {
  52         .init_name = "fallback device",
  53         .coherent_dma_mask = ISA_DMA_BIT_MASK,
  54         .dma_mask = &x86_dma_fallback_dev.coherent_dma_mask,
  55 };
  56 EXPORT_SYMBOL(x86_dma_fallback_dev);
  57
  58 /* Number of entries preallocated for DMA-API debugging */
  59 #define PREALLOC_DMA_DEBUG_ENTRIES       32768
  60
  61 int dma_set_mask(struct device *dev, u64 mask)
  62 {
  63         if (!dev->dma_mask || !dma_supported(dev, mask))
  64                 return -EIO;
  65
  66         *dev->dma_mask = mask;
  67
  68         return 0;
  69 }
  70 EXPORT_SYMBOL(dma_set_mask);
  71
  72 void __init pci_iommu_alloc(void)
  73 {
  74         struct iommu_table_entry *p;
  75
  76         sort_iommu_table(__iommu_table, __iommu_table_end);
  77         check_iommu_entries(__iommu_table, __iommu_table_end);
  78
  79         for (p = __iommu_table; p < __iommu_table_end; p++) {
  80                 if (p && p->detect && p->detect() > 0) {
  81                         p->flags |= IOMMU_DETECTED;
  82                         if (p->early_init)
  83                                 p->early_init();
  84                         if (p->flags & IOMMU_FINISH_IF_DETECTED)
  85                                 break;
  86                 }
  87         }
  88 }
  89 void *dma_generic_alloc_coherent(struct device *dev, size_t size,
  90                                  dma_addr_t *dma_addr, gfp_t flag)
  91 {
  92         unsigned long dma_mask;
  93         struct page *page;
  94         dma_addr_t addr;
  95
  96         dma_mask = dma_alloc_coherent_mask(dev, flag);
  97
  98         flag |= __GFP_ZERO;
  99 again:
 100         page = alloc_pages_node(dev_to_node(dev), flag, get_order(size));
 101         if (!page)
 102                 return NULL;
 103
 104         addr = page_to_phys(page);
 105         if (addr + size > dma_mask) {
 106                 __free_pages(page, get_order(size));
 107
 108                 if (dma_mask < DMA_BIT_MASK(32) && !(flag & GFP_DMA)) {
 109                         flag = (flag & ~GFP_DMA32) | GFP_DMA;
 110                         goto again;
 111                 }
 112
 113                 return NULL;
 114         }
 115
 116         *dma_addr = addr;
 117         return page_address(page);
 118 }
 119
 120 /*
 121  * See <Documentation/x86_64/boot-options.txt> for the iommu kernel parameter
 122  * documentation.
 123  */
 124 static __init int iommu_setup(char *p)
 125 {
 126         iommu_merge = 1;
 127
 128         if (!p)
 129                 return -EINVAL;
 130
 131         while (*p) {
 132                 if (!strncmp(p, "off", 3))
 133                         no_iommu = 1;
 134                 /* gart_parse_options has more force support */
 135                 if (!strncmp(p, "force", 5))
 136                         force_iommu = 1;
 137                 if (!strncmp(p, "noforce", 7)) {
 138                         iommu_merge = 0;
 139                         force_iommu = 0;
 140                 }
 141
 142                 if (!strncmp(p, "biomerge", 8)) {
 143                         iommu_merge = 1;
 144                         force_iommu = 1;
 145                 }
 146                 if (!strncmp(p, "panic", 5))
 147                         panic_on_overflow = 1;
 148                 if (!strncmp(p, "nopanic", 7))
 149                         panic_on_overflow = 0;
 150                 if (!strncmp(p, "merge", 5)) {
 151                         iommu_merge = 1;
 152                         force_iommu = 1;
 153                 }
 154                 if (!strncmp(p, "nomerge", 7))
 155                         iommu_merge = 0;
 156                 if (!strncmp(p, "forcesac", 8))
 157                         iommu_sac_force = 1;
 158                 if (!strncmp(p, "allowdac", 8))
 159                         forbid_dac = 0;
 160                 if (!strncmp(p, "nodac", 5))
 161                         forbid_dac = 1;
 162                 if (!strncmp(p, "usedac", 6)) {
 163                         forbid_dac = -1;
 164                         return 1;
 165                 }
 166 #ifdef CONFIG_SWIOTLB
 167                 if (!strncmp(p, "soft", 4))
 168                         swiotlb = 1;
 169 #endif
 170                 if (!strncmp(p, "pt", 2))
 171                         iommu_pass_through = 1;
 172
 173                 gart_parse_options(p);
 174
 175 #ifdef CONFIG_CALGARY_IOMMU
 176                 if (!strncmp(p, "calgary", 7))
 177                         use_calgary = 1;
 178 #endif /* CONFIG_CALGARY_IOMMU */
 179
 180                 p += strcspn(p, ",");
 181                 if (*p == ',')
 182                         ++p;
 183         }
 184         return 0;
 185 }
 186 early_param("iommu", iommu_setup);
 187
 188 int dma_supported(struct device *dev, u64 mask)
 189 {
 190         struct dma_map_ops *ops = get_dma_ops(dev);
 191
 192 #ifdef CONFIG_PCI
 193         if (mask > 0xffffffff && forbid_dac > 0) {
 194                 dev_info(dev, "PCI: Disallowing DAC for device\n");
 195                 return 0;
 196         }
 197 #endif
 198
 199         if (ops->dma_supported)
 200                 return ops->dma_supported(dev, mask);
 201
 202         /* Copied from i386. Doesn't make much sense, because it will
 203            only work for pci_alloc_coherent.
 204            The caller just has to use GFP_DMA in this case. */
 205         if (mask < DMA_BIT_MASK(24))
 206                 return 0;
 207
 208         /* Tell the device to use SAC when IOMMU force is on.  This
 209            allows the driver to use cheaper accesses in some cases.
 210
 211            Problem with this is that if we overflow the IOMMU area and
 212            return DAC as fallback address the device may not handle it
 213            correctly.
 214
 215            As a special case some controllers have a 39bit address
 216            mode that is as efficient as 32bit (aic79xx). Don't force
 217            SAC for these.  Assume all masks <= 40 bits are of this
 218            type. Normally this doesn't make any difference, but gives
 219            more gentle handling of IOMMU overflow. */
 220         if (iommu_sac_force && (mask >= DMA_BIT_MASK(40))) {
 221                 dev_info(dev, "Force SAC with mask %Lx\n", mask);
 222                 return 0;
 223         }
 224
 225         return 1;
 226 }
 227 EXPORT_SYMBOL(dma_supported);
 228
 229 static int __init pci_iommu_init(void)
 230 {
 231         struct iommu_table_entry *p;
 232         dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
 233
 234 #ifdef CONFIG_PCI
 235         dma_debug_add_bus(&pci_bus_type);
 236 #endif
 237         x86_init.iommu.iommu_init();
 238
 239         for (p = __iommu_table; p < __iommu_table_end; p++) {
 240                 if (p && (p->flags & IOMMU_DETECTED) && p->late_init)
 241                         p->late_init();
 242         }
 243
 244         return 0;
 245 }
 246 /* Must execute after PCI subsystem */
 247 rootfs_initcall(pci_iommu_init);
 248
 249 #ifdef CONFIG_PCI
 250 /* Many VIA bridges seem to corrupt data for DAC. Disable it here */
 251
 252 static __devinit void via_no_dac(struct pci_dev *dev)
 253 {
 254         if ((dev->class >> 8) == PCI_CLASS_BRIDGE_PCI && forbid_dac == 0) {
 255                 dev_info(&dev->dev, "disabling DAC on VIA PCI bridge\n");
 256                 forbid_dac = 1;
 257         }
 258 }
 259 DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_VIA, PCI_ANY_ID, via_no_dac);
 260 #endif