linux/Documentation/bus-virt-phys-mapping.txt
<<
>>
Prefs
   1==========================================================
   2How to access I/O mapped memory from within device drivers
   3==========================================================
   4
   5:Author: Linus
   6
   7.. warning::
   8
   9        The virt_to_bus() and bus_to_virt() functions have been
  10        superseded by the functionality provided by the PCI DMA interface
  11        (see Documentation/DMA-API-HOWTO.txt).  They continue
  12        to be documented below for historical purposes, but new code
  13        must not use them. --davidm 00/12/12
  14
  15::
  16
  17  [ This is a mail message in response to a query on IO mapping, thus the
  18    strange format for a "document" ]
  19
  20The AHA-1542 is a bus-master device, and your patch makes the driver give the
  21controller the physical address of the buffers, which is correct on x86
  22(because all bus master devices see the physical memory mappings directly). 
  23
  24However, on many setups, there are actually **three** different ways of looking
  25at memory addresses, and in this case we actually want the third, the
  26so-called "bus address". 
  27
  28Essentially, the three ways of addressing memory are (this is "real memory",
  29that is, normal RAM--see later about other details): 
  30
  31 - CPU untranslated.  This is the "physical" address.  Physical address 
  32   0 is what the CPU sees when it drives zeroes on the memory bus.
  33
  34 - CPU translated address. This is the "virtual" address, and is 
  35   completely internal to the CPU itself with the CPU doing the appropriate
  36   translations into "CPU untranslated". 
  37
  38 - bus address. This is the address of memory as seen by OTHER devices, 
  39   not the CPU. Now, in theory there could be many different bus 
  40   addresses, with each device seeing memory in some device-specific way, but
  41   happily most hardware designers aren't actually actively trying to make
  42   things any more complex than necessary, so you can assume that all 
  43   external hardware sees the memory the same way. 
  44
  45Now, on normal PCs the bus address is exactly the same as the physical
  46address, and things are very simple indeed. However, they are that simple
  47because the memory and the devices share the same address space, and that is
  48not generally necessarily true on other PCI/ISA setups. 
  49
  50Now, just as an example, on the PReP (PowerPC Reference Platform), the 
  51CPU sees a memory map something like this (this is from memory)::
  52
  53        0-2 GB          "real memory"
  54        2 GB-3 GB       "system IO" (inb/out and similar accesses on x86)
  55        3 GB-4 GB       "IO memory" (shared memory over the IO bus)
  56
  57Now, that looks simple enough. However, when you look at the same thing from
  58the viewpoint of the devices, you have the reverse, and the physical memory
  59address 0 actually shows up as address 2 GB for any IO master.
  60
  61So when the CPU wants any bus master to write to physical memory 0, it 
  62has to give the master address 0x80000000 as the memory address.
  63
  64So, for example, depending on how the kernel is actually mapped on the 
  65PPC, you can end up with a setup like this::
  66
  67 physical address:      0
  68 virtual address:       0xC0000000
  69 bus address:           0x80000000
  70
  71where all the addresses actually point to the same thing.  It's just seen 
  72through different translations..
  73
  74Similarly, on the Alpha, the normal translation is::
  75
  76 physical address:      0
  77 virtual address:       0xfffffc0000000000
  78 bus address:           0x40000000
  79
  80(but there are also Alphas where the physical address and the bus address
  81are the same). 
  82
  83Anyway, the way to look up all these translations, you do::
  84
  85        #include <asm/io.h>
  86
  87        phys_addr = virt_to_phys(virt_addr);
  88        virt_addr = phys_to_virt(phys_addr);
  89         bus_addr = virt_to_bus(virt_addr);
  90        virt_addr = bus_to_virt(bus_addr);
  91
  92Now, when do you need these?
  93
  94You want the **virtual** address when you are actually going to access that
  95pointer from the kernel. So you can have something like this::
  96
  97        /*
  98         * this is the hardware "mailbox" we use to communicate with
  99         * the controller. The controller sees this directly.
 100         */
 101        struct mailbox {
 102                __u32 status;
 103                __u32 bufstart;
 104                __u32 buflen;
 105                ..
 106        } mbox;
 107
 108                unsigned char * retbuffer;
 109
 110                /* get the address from the controller */
 111                retbuffer = bus_to_virt(mbox.bufstart);
 112                switch (retbuffer[0]) {
 113                        case STATUS_OK:
 114                                ...
 115
 116on the other hand, you want the bus address when you have a buffer that 
 117you want to give to the controller::
 118
 119        /* ask the controller to read the sense status into "sense_buffer" */
 120        mbox.bufstart = virt_to_bus(&sense_buffer);
 121        mbox.buflen = sizeof(sense_buffer);
 122        mbox.status = 0;
 123        notify_controller(&mbox);
 124
 125And you generally **never** want to use the physical address, because you can't
 126use that from the CPU (the CPU only uses translated virtual addresses), and
 127you can't use it from the bus master. 
 128
 129So why do we care about the physical address at all? We do need the physical
 130address in some cases, it's just not very often in normal code.  The physical
 131address is needed if you use memory mappings, for example, because the
 132"remap_pfn_range()" mm function wants the physical address of the memory to
 133be remapped as measured in units of pages, a.k.a. the pfn (the memory
 134management layer doesn't know about devices outside the CPU, so it
 135shouldn't need to know about "bus addresses" etc).
 136
 137.. note::
 138
 139        The above is only one part of the whole equation. The above
 140        only talks about "real memory", that is, CPU memory (RAM).
 141
 142There is a completely different type of memory too, and that's the "shared
 143memory" on the PCI or ISA bus. That's generally not RAM (although in the case
 144of a video graphics card it can be normal DRAM that is just used for a frame
 145buffer), but can be things like a packet buffer in a network card etc. 
 146
 147This memory is called "PCI memory" or "shared memory" or "IO memory" or
 148whatever, and there is only one way to access it: the readb/writeb and
 149related functions. You should never take the address of such memory, because
 150there is really nothing you can do with such an address: it's not
 151conceptually in the same memory space as "real memory" at all, so you cannot
 152just dereference a pointer. (Sadly, on x86 it **is** in the same memory space,
 153so on x86 it actually works to just deference a pointer, but it's not
 154portable). 
 155
 156For such memory, you can do things like:
 157
 158 - reading::
 159
 160        /*
 161         * read first 32 bits from ISA memory at 0xC0000, aka
 162         * C000:0000 in DOS terms
 163         */
 164        unsigned int signature = isa_readl(0xC0000);
 165
 166 - remapping and writing::
 167
 168        /*
 169         * remap framebuffer PCI memory area at 0xFC000000,
 170         * size 1MB, so that we can access it: We can directly
 171         * access only the 640k-1MB area, so anything else
 172         * has to be remapped.
 173         */
 174        void __iomem *baseptr = ioremap(0xFC000000, 1024*1024);
 175
 176        /* write a 'A' to the offset 10 of the area */
 177        writeb('A',baseptr+10);
 178
 179        /* unmap when we unload the driver */
 180        iounmap(baseptr);
 181
 182 - copying and clearing::
 183
 184        /* get the 6-byte Ethernet address at ISA address E000:0040 */
 185        memcpy_fromio(kernel_buffer, 0xE0040, 6);
 186        /* write a packet to the driver */
 187        memcpy_toio(0xE1000, skb->data, skb->len);
 188        /* clear the frame buffer */
 189        memset_io(0xA0000, 0, 0x10000);
 190
 191OK, that just about covers the basics of accessing IO portably.  Questions?
 192Comments? You may think that all the above is overly complex, but one day you
 193might find yourself with a 500 MHz Alpha in front of you, and then you'll be
 194happy that your driver works ;)
 195
 196Note that kernel versions 2.0.x (and earlier) mistakenly called the
 197ioremap() function "vremap()".  ioremap() is the proper name, but I
 198didn't think straight when I wrote it originally.  People who have to
 199support both can do something like::
 200 
 201        /* support old naming silliness */
 202        #if LINUX_VERSION_CODE < 0x020100
 203        #define ioremap vremap
 204        #define iounmap vfree                                                     
 205        #endif
 206 
 207at the top of their source files, and then they can use the right names
 208even on 2.0.x systems. 
 209
 210And the above sounds worse than it really is.  Most real drivers really
 211don't do all that complex things (or rather: the complexity is not so
 212much in the actual IO accesses as in error handling and timeouts etc). 
 213It's generally not hard to fix drivers, and in many cases the code
 214actually looks better afterwards::
 215
 216        unsigned long signature = *(unsigned int *) 0xC0000;
 217                vs
 218        unsigned long signature = readl(0xC0000);
 219
 220I think the second version actually is more readable, no?
 221