Documentation/x86/pat.txt

   1
   2 PAT (Page Attribute Table)
   3
   4 x86 Page Attribute Table (PAT) allows for setting the memory attribute at the
   5 page level granularity. PAT is complementary to the MTRR settings which allows
   6 for setting of memory types over physical address ranges. However, PAT is
   7 more flexible than MTRR due to its capability to set attributes at page level
   8 and also due to the fact that there are no hardware limitations on number of
   9 such attribute settings allowed. Added flexibility comes with guidelines for
  10 not having memory type aliasing for the same physical memory with multiple
  11 virtual addresses.
  12
  13 PAT allows for different types of memory attributes. The most commonly used
  14 ones that will be supported at this time are Write-back, Uncached,
  15 Write-combined and Uncached Minus.
  16
  17
  18 PAT APIs
  19 --------
  20
  21 There are many different APIs in the kernel that allows setting of memory
  22 attributes at the page level. In order to avoid aliasing, these interfaces
  23 should be used thoughtfully. Below is a table of interfaces available,
  24 their intended usage and their memory attribute relationships. Internally,
  25 these APIs use a reserve_memtype()/free_memtype() interface on the physical
  26 address range to avoid any aliasing.
  27
  28
  29 -------------------------------------------------------------------
  30 API                    |    RAM   |  ACPI,...  |  Reserved/Holes  |
  31 -----------------------|----------|------------|------------------|
  32                        |          |            |                  |
  33 ioremap                |    --    |    UC-     |       UC-        |
  34                        |          |            |                  |
  35 ioremap_cache          |    --    |    WB      |       WB         |
  36                        |          |            |                  |
  37 ioremap_nocache        |    --    |    UC-     |       UC-        |
  38                        |          |            |                  |
  39 ioremap_wc             |    --    |    --      |       WC         |
  40                        |          |            |                  |
  41 set_memory_uc          |    UC-   |    --      |       --         |
  42  set_memory_wb         |          |            |                  |
  43                        |          |            |                  |
  44 set_memory_wc          |    WC    |    --      |       --         |
  45  set_memory_wb         |          |            |                  |
  46                        |          |            |                  |
  47 pci sysfs resource     |    --    |    --      |       UC-        |
  48                        |          |            |                  |
  49 pci sysfs resource_wc  |    --    |    --      |       WC         |
  50  is IORESOURCE_PREFETCH|          |            |                  |
  51                        |          |            |                  |
  52 pci proc               |    --    |    --      |       UC-        |
  53  !PCIIOC_WRITE_COMBINE |          |            |                  |
  54                        |          |            |                  |
  55 pci proc               |    --    |    --      |       WC         |
  56  PCIIOC_WRITE_COMBINE  |          |            |                  |
  57                        |          |            |                  |
  58 /dev/mem               |    --    |  WB/WC/UC- |    WB/WC/UC-     |
  59  read-write            |          |            |                  |
  60                        |          |            |                  |
  61 /dev/mem               |    --    |    UC-     |       UC-        |
  62  mmap SYNC flag        |          |            |                  |
  63                        |          |            |                  |
  64 /dev/mem               |    --    |  WB/WC/UC- |    WB/WC/UC-     |
  65  mmap !SYNC flag       |          |(from exist-|  (from exist-    |
  66  and                   |          |  ing alias)|    ing alias)    |
  67  any alias to this area|          |            |                  |
  68                        |          |            |                  |
  69 /dev/mem               |    --    |    WB      |       WB         |
  70  mmap !SYNC flag       |          |            |                  |
  71  no alias to this area |          |            |                  |
  72  and                   |          |            |                  |
  73  MTRR says WB          |          |            |                  |
  74                        |          |            |                  |
  75 /dev/mem               |    --    |    --      |       UC-        |
  76  mmap !SYNC flag       |          |            |                  |
  77  no alias to this area |          |            |                  |
  78  and                   |          |            |                  |
  79  MTRR says !WB         |          |            |                  |
  80                        |          |            |                  |
  81 -------------------------------------------------------------------
  82
  83 Advanced APIs for drivers
  84 -------------------------
  85 A. Exporting pages to users with remap_pfn_range, io_remap_pfn_range,
  86 vm_insert_pfn
  87
  88 Drivers wanting to export some pages to userspace do it by using mmap
  89 interface and a combination of
  90 1) pgprot_noncached()
  91 2) io_remap_pfn_range() or remap_pfn_range() or vm_insert_pfn()
  92
  93 With PAT support, a new API pgprot_writecombine is being added. So, drivers can
  94 continue to use the above sequence, with either pgprot_noncached() or
  95 pgprot_writecombine() in step 1, followed by step 2.
  96
  97 In addition, step 2 internally tracks the region as UC or WC in memtype
  98 list in order to ensure no conflicting mapping.
  99
 100 Note that this set of APIs only works with IO (non RAM) regions. If driver
 101 wants to export a RAM region, it has to do set_memory_uc() or set_memory_wc()
 102 as step 0 above and also track the usage of those pages and use set_memory_wb()
 103 before the page is freed to free pool.
 104
 105
 106
 107 Notes:
 108
 109 -- in the above table mean "Not suggested usage for the API". Some of the --'s
 110 are strictly enforced by the kernel. Some others are not really enforced
 111 today, but may be enforced in future.
 112
 113 For ioremap and pci access through /sys or /proc - The actual type returned
 114 can be more restrictive, in case of any existing aliasing for that address.
 115 For example: If there is an existing uncached mapping, a new ioremap_wc can
 116 return uncached mapping in place of write-combine requested.
 117
 118 set_memory_[uc|wc] and set_memory_wb should be used in pairs, where driver will
 119 first make a region uc or wc and switch it back to wb after use.
 120
 121 Over time writes to /proc/mtrr will be deprecated in favor of using PAT based
 122 interfaces. Users writing to /proc/mtrr are suggested to use above interfaces.
 123
 124 Drivers should use ioremap_[uc|wc] to access PCI BARs with [uc|wc] access
 125 types.
 126
 127 Drivers should use set_memory_[uc|wc] to set access type for RAM ranges.
 128
 129
 130 PAT debugging
 131 -------------
 132
 133 With CONFIG_DEBUG_FS enabled, PAT memtype list can be examined by
 134
 135 # mount -t debugfs debugfs /sys/kernel/debug
 136 # cat /sys/kernel/debug/x86/pat_memtype_list
 137 PAT memtype list:
 138 uncached-minus @ 0x7fadf000-0x7fae0000
 139 uncached-minus @ 0x7fb19000-0x7fb1a000
 140 uncached-minus @ 0x7fb1a000-0x7fb1b000
 141 uncached-minus @ 0x7fb1b000-0x7fb1c000
 142 uncached-minus @ 0x7fb1c000-0x7fb1d000
 143 uncached-minus @ 0x7fb1d000-0x7fb1e000
 144 uncached-minus @ 0x7fb1e000-0x7fb25000
 145 uncached-minus @ 0x7fb25000-0x7fb26000
 146 uncached-minus @ 0x7fb26000-0x7fb27000
 147 uncached-minus @ 0x7fb27000-0x7fb28000
 148 uncached-minus @ 0x7fb28000-0x7fb2e000
 149 uncached-minus @ 0x7fb2e000-0x7fb2f000
 150 uncached-minus @ 0x7fb2f000-0x7fb30000
 151 uncached-minus @ 0x7fb31000-0x7fb32000
 152 uncached-minus @ 0x80000000-0x90000000
 153
 154 This list shows physical address ranges and various PAT settings used to
 155 access those physical address ranges.
 156
 157 Another, more verbose way of getting PAT related debug messages is with
 158 "debugpat" boot parameter. With this parameter, various debug messages are
 159 printed to dmesg log.
 160