1                     Dynamic DMA mapping Guide
   2                     =========================
   4                 David S. Miller <>
   5                 Richard Henderson <>
   6                  Jakub Jelinek <>
   8This is a guide to device driver writers on how to use the DMA API
   9with example pseudo-code.  For a concise description of the API, see
  12                       CPU and DMA addresses
  14There are several kinds of addresses involved in the DMA API, and it's
  15important to understand the differences.
  17The kernel normally uses virtual addresses.  Any address returned by
  18kmalloc(), vmalloc(), and similar interfaces is a virtual address and can
  19be stored in a "void *".
  21The virtual memory system (TLB, page tables, etc.) translates virtual
  22addresses to CPU physical addresses, which are stored as "phys_addr_t" or
  23"resource_size_t".  The kernel manages device resources like registers as
  24physical addresses.  These are the addresses in /proc/iomem.  The physical
  25address is not directly useful to a driver; it must use ioremap() to map
  26the space and produce a virtual address.
  28I/O devices use a third kind of address: a "bus address" or "DMA address".
  29If a device has registers at an MMIO address, or if it performs DMA to read
  30or write system memory, the addresses used by the device are bus addresses.
  31In some systems, bus addresses are identical to CPU physical addresses, but
  32in general they are not.  IOMMUs and host bridges can produce arbitrary
  33mappings between physical and bus addresses.
  35Here's a picture and some examples:
  37               CPU                  CPU                  Bus
  38             Virtual              Physical             Address
  39             Address              Address               Space
  40              Space                Space
  42            +-------+             +------+             +------+
  43            |       |             |MMIO  |   Offset    |      |
  44            |       |  Virtual    |Space |   applied   |      |
  45          C +-------+ --------> B +------+ ----------> +------+ A
  46            |       |  mapping    |      |   by host   |      |
  47  +-----+   |       |             |      |   bridge    |      |   +--------+
  48  |     |   |       |             +------+             |      |   |        |
  49  | CPU |   |       |             | RAM  |             |      |   | Device |
  50  |     |   |       |             |      |             |      |   |        |
  51  +-----+   +-------+             +------+             +------+   +--------+
  52            |       |  Virtual    |Buffer|   Mapping   |      |
  53          X +-------+ --------> Y +------+ <---------- +------+ Z
  54            |       |  mapping    | RAM  |   by IOMMU
  55            |       |             |      |
  56            |       |             |      |
  57            +-------+             +------+
  59During the enumeration process, the kernel learns about I/O devices and
  60their MMIO space and the host bridges that connect them to the system.  For
  61example, if a PCI device has a BAR, the kernel reads the bus address (A)
  62from the BAR and converts it to a CPU physical address (B).  The address B
  63is stored in a struct resource and usually exposed via /proc/iomem.  When a
  64driver claims a device, it typically uses ioremap() to map physical address
  65B at a virtual address (C).  It can then use, e.g., ioread32(C), to access
  66the device registers at bus address A.
  68If the device supports DMA, the driver sets up a buffer using kmalloc() or
  69a similar interface, which returns a virtual address (X).  The virtual
  70memory system maps X to a physical address (Y) in system RAM.  The driver
  71can use virtual address X to access the buffer, but the device itself
  72cannot because DMA doesn't go through the CPU virtual memory system.
  74In some simple systems, the device can do DMA directly to physical address
  75Y.  But in many others, there is IOMMU hardware that translates bus
  76addresses to physical addresses, e.g., it translates Z to Y.  This is part
  77of the reason for the DMA API: the driver can give a virtual address X to
  78an interface like dma_map_single(), which sets up any required IOMMU
  79mapping and returns the bus address Z.  The driver then tells the device to
  80do DMA to Z, and the IOMMU maps it to the buffer at address Y in system
  83So that Linux can use the dynamic DMA mapping, it needs some help from the
  84drivers, namely it has to take into account that DMA addresses should be
  85mapped only for the time they are actually used and unmapped after the DMA
  88The following API will work of course even on platforms where no such
  89hardware exists.
  91Note that the DMA API works with any bus independent of the underlying
  92microprocessor architecture. You should use the DMA API rather than the
  93bus-specific DMA API, i.e., use the dma_map_*() interfaces rather than the
  94pci_map_*() interfaces.
  96First of all, you should make sure
  98#include <linux/dma-mapping.h>
 100is in your driver, which provides the definition of dma_addr_t.  This type
 101can hold any valid DMA or bus address for the platform and should be used
 102everywhere you hold a DMA address returned from the DMA mapping functions.
 104                         What memory is DMA'able?
 106The first piece of information you must know is what kernel memory can
 107be used with the DMA mapping facilities.  There has been an unwritten
 108set of rules regarding this, and this text is an attempt to finally
 109write them down.
 111If you acquired your memory via the page allocator
 112(i.e. __get_free_page*()) or the generic memory allocators
 113(i.e. kmalloc() or kmem_cache_alloc()) then you may DMA to/from
 114that memory using the addresses returned from those routines.
 116This means specifically that you may _not_ use the memory/addresses
 117returned from vmalloc() for DMA.  It is possible to DMA to the
 118_underlying_ memory mapped into a vmalloc() area, but this requires
 119walking page tables to get the physical addresses, and then
 120translating each of those pages back to a kernel address using
 121something like __va().  [ EDIT: Update this when we integrate
 122Gerd Knorr's generic code which does this. ]
 124This rule also means that you may use neither kernel image addresses
 125(items in data/text/bss segments), nor module image addresses, nor
 126stack addresses for DMA.  These could all be mapped somewhere entirely
 127different than the rest of physical memory.  Even if those classes of
 128memory could physically work with DMA, you'd need to ensure the I/O
 129buffers were cacheline-aligned.  Without that, you'd see cacheline
 130sharing problems (data corruption) on CPUs with DMA-incoherent caches.
 131(The CPU could write to one word, DMA would write to a different one
 132in the same cache line, and one of them could be overwritten.)
 134Also, this means that you cannot take the return of a kmap()
 135call and DMA to/from that.  This is similar to vmalloc().
 137What about block I/O and networking buffers?  The block I/O and
 138networking subsystems make sure that the buffers they use are valid
 139for you to DMA from/to.
 141                        DMA addressing limitations
 143Does your device have any DMA addressing limitations?  For example, is
 144your device only capable of driving the low order 24-bits of address?
 145If so, you need to inform the kernel of this fact.
 147By default, the kernel assumes that your device can address the full
 14832-bits.  For a 64-bit capable device, this needs to be increased.
 149And for a device with limitations, as discussed in the previous
 150paragraph, it needs to be decreased.
 152Special note about PCI: PCI-X specification requires PCI-X devices to
 153support 64-bit addressing (DAC) for all transactions.  And at least
 154one platform (SGI SN2) requires 64-bit consistent allocations to
 155operate correctly when the IO bus is in PCI-X mode.
 157For correct operation, you must interrogate the kernel in your device
 158probe routine to see if the DMA controller on the machine can properly
 159support the DMA addressing limitation your device has.  It is good
 160style to do this even if your device holds the default setting,
 161because this shows that you did think about these issues wrt. your
 164The query is performed via a call to dma_set_mask_and_coherent():
 166        int dma_set_mask_and_coherent(struct device *dev, u64 mask);
 168which will query the mask for both streaming and coherent APIs together.
 169If you have some special requirements, then the following two separate
 170queries can be used instead:
 172        The query for streaming mappings is performed via a call to
 173        dma_set_mask():
 175                int dma_set_mask(struct device *dev, u64 mask);
 177        The query for consistent allocations is performed via a call
 178        to dma_set_coherent_mask():
 180                int dma_set_coherent_mask(struct device *dev, u64 mask);
 182Here, dev is a pointer to the device struct of your device, and mask
 183is a bit mask describing which bits of an address your device
 184supports.  It returns zero if your card can perform DMA properly on
 185the machine given the address mask you provided.  In general, the
 186device struct of your device is embedded in the bus-specific device
 187struct of your device.  For example, &pdev->dev is a pointer to the
 188device struct of a PCI device (pdev is a pointer to the PCI device
 189struct of your device).
 191If it returns non-zero, your device cannot perform DMA properly on
 192this platform, and attempting to do so will result in undefined
 193behavior.  You must either use a different mask, or not use DMA.
 195This means that in the failure case, you have three options:
 1971) Use another DMA mask, if possible (see below).
 1982) Use some non-DMA mode for data transfer, if possible.
 1993) Ignore this device and do not initialize it.
 201It is recommended that your driver print a kernel KERN_WARNING message
 202when you end up performing either #2 or #3.  In this manner, if a user
 203of your driver reports that performance is bad or that the device is not
 204even detected, you can ask them for the kernel messages to find out
 205exactly why.
 207The standard 32-bit addressing device would do something like this:
 209        if (dma_set_mask_and_coherent(dev, DMA_BIT_MASK(32))) {
 210                dev_warn(dev, "mydev: No suitable DMA available\n");
 211                goto ignore_this_device;
 212        }
 214Another common scenario is a 64-bit capable device.  The approach here
 215is to try for 64-bit addressing, but back down to a 32-bit mask that
 216should not fail.  The kernel may fail the 64-bit mask not because the
 217platform is not capable of 64-bit addressing.  Rather, it may fail in
 218this case simply because 32-bit addressing is done more efficiently
 219than 64-bit addressing.  For example, Sparc64 PCI SAC addressing is
 220more efficient than DAC addressing.
 222Here is how you would handle a 64-bit capable device which can drive
 223all 64-bits when accessing streaming DMA:
 225        int using_dac;
 227        if (!dma_set_mask(dev, DMA_BIT_MASK(64))) {
 228                using_dac = 1;
 229        } else if (!dma_set_mask(dev, DMA_BIT_MASK(32))) {
 230                using_dac = 0;
 231        } else {
 232                dev_warn(dev, "mydev: No suitable DMA available\n");
 233                goto ignore_this_device;
 234        }
 236If a card is capable of using 64-bit consistent allocations as well,
 237the case would look like this:
 239        int using_dac, consistent_using_dac;
 241        if (!dma_set_mask_and_coherent(dev, DMA_BIT_MASK(64))) {
 242                using_dac = 1;
 243                consistent_using_dac = 1;
 244        } else if (!dma_set_mask_and_coherent(dev, DMA_BIT_MASK(32))) {
 245                using_dac = 0;
 246                consistent_using_dac = 0;
 247        } else {
 248                dev_warn(dev, "mydev: No suitable DMA available\n");
 249                goto ignore_this_device;
 250        }
 252The coherent mask will always be able to set the same or a smaller mask as
 253the streaming mask. However for the rare case that a device driver only
 254uses consistent allocations, one would have to check the return value from
 257Finally, if your device can only drive the low 24-bits of
 258address you might do something like:
 260        if (dma_set_mask(dev, DMA_BIT_MASK(24))) {
 261                dev_warn(dev, "mydev: 24-bit DMA addressing not available\n");
 262                goto ignore_this_device;
 263        }
 265When dma_set_mask() or dma_set_mask_and_coherent() is successful, and
 266returns zero, the kernel saves away this mask you have provided.  The
 267kernel will use this information later when you make DMA mappings.
 269There is a case which we are aware of at this time, which is worth
 270mentioning in this documentation.  If your device supports multiple
 271functions (for example a sound card provides playback and record
 272functions) and the various different functions have _different_
 273DMA addressing limitations, you may wish to probe each mask and
 274only provide the functionality which the machine can handle.  It
 275is important that the last call to dma_set_mask() be for the
 276most specific mask.
 278Here is pseudo-code showing how this might be done:
 280        #define PLAYBACK_ADDRESS_BITS   DMA_BIT_MASK(32)
 281        #define RECORD_ADDRESS_BITS     DMA_BIT_MASK(24)
 283        struct my_sound_card *card;
 284        struct device *dev;
 286        ...
 287        if (!dma_set_mask(dev, PLAYBACK_ADDRESS_BITS)) {
 288                card->playback_enabled = 1;
 289        } else {
 290                card->playback_enabled = 0;
 291                dev_warn(dev, "%s: Playback disabled due to DMA limitations\n",
 292                       card->name);
 293        }
 294        if (!dma_set_mask(dev, RECORD_ADDRESS_BITS)) {
 295                card->record_enabled = 1;
 296        } else {
 297                card->record_enabled = 0;
 298                dev_warn(dev, "%s: Record disabled due to DMA limitations\n",
 299                       card->name);
 300        }
 302A sound card was used as an example here because this genre of PCI
 303devices seems to be littered with ISA chips given a PCI front end,
 304and thus retaining the 16MB DMA addressing limitations of ISA.
 306                        Types of DMA mappings
 308There are two types of DMA mappings:
 310- Consistent DMA mappings which are usually mapped at driver
 311  initialization, unmapped at the end and for which the hardware should
 312  guarantee that the device and the CPU can access the data
 313  in parallel and will see updates made by each other without any
 314  explicit software flushing.
 316  Think of "consistent" as "synchronous" or "coherent".
 318  The current default is to return consistent memory in the low 32
 319  bits of the bus space.  However, for future compatibility you should
 320  set the consistent mask even if this default is fine for your
 321  driver.
 323  Good examples of what to use consistent mappings for are:
 325        - Network card DMA ring descriptors.
 326        - SCSI adapter mailbox command data structures.
 327        - Device firmware microcode executed out of
 328          main memory.
 330  The invariant these examples all require is that any CPU store
 331  to memory is immediately visible to the device, and vice
 332  versa.  Consistent mappings guarantee this.
 334  IMPORTANT: Consistent DMA memory does not preclude the usage of
 335             proper memory barriers.  The CPU may reorder stores to
 336             consistent memory just as it may normal memory.  Example:
 337             if it is important for the device to see the first word
 338             of a descriptor updated before the second, you must do
 339             something like:
 341                desc->word0 = address;
 342                wmb();
 343                desc->word1 = DESC_VALID;
 345             in order to get correct behavior on all platforms.
 347             Also, on some platforms your driver may need to flush CPU write
 348             buffers in much the same way as it needs to flush write buffers
 349             found in PCI bridges (such as by reading a register's value
 350             after writing it).
 352- Streaming DMA mappings which are usually mapped for one DMA
 353  transfer, unmapped right after it (unless you use dma_sync_* below)
 354  and for which hardware can optimize for sequential accesses.
 356  This of "streaming" as "asynchronous" or "outside the coherency
 357  domain".
 359  Good examples of what to use streaming mappings for are:
 361        - Networking buffers transmitted/received by a device.
 362        - Filesystem buffers written/read by a SCSI device.
 364  The interfaces for using this type of mapping were designed in
 365  such a way that an implementation can make whatever performance
 366  optimizations the hardware allows.  To this end, when using
 367  such mappings you must be explicit about what you want to happen.
 369Neither type of DMA mapping has alignment restrictions that come from
 370the underlying bus, although some devices may have such restrictions.
 371Also, systems with caches that aren't DMA-coherent will work better
 372when the underlying buffers don't share cache lines with other data.
 375                 Using Consistent DMA mappings.
 377To allocate and map large (PAGE_SIZE or so) consistent DMA regions,
 378you should do:
 380        dma_addr_t dma_handle;
 382        cpu_addr = dma_alloc_coherent(dev, size, &dma_handle, gfp);
 384where device is a struct device *. This may be called in interrupt
 385context with the GFP_ATOMIC flag.
 387Size is the length of the region you want to allocate, in bytes.
 389This routine will allocate RAM for that region, so it acts similarly to
 390__get_free_pages() (but takes size instead of a page order).  If your
 391driver needs regions sized smaller than a page, you may prefer using
 392the dma_pool interface, described below.
 394The consistent DMA mapping interfaces, for non-NULL dev, will by
 395default return a DMA address which is 32-bit addressable.  Even if the
 396device indicates (via DMA mask) that it may address the upper 32-bits,
 397consistent allocation will only return > 32-bit addresses for DMA if
 398the consistent DMA mask has been explicitly changed via
 399dma_set_coherent_mask().  This is true of the dma_pool interface as
 402dma_alloc_coherent() returns two values: the virtual address which you
 403can use to access it from the CPU and dma_handle which you pass to the
 406The CPU virtual address and the DMA bus address are both
 407guaranteed to be aligned to the smallest PAGE_SIZE order which
 408is greater than or equal to the requested size.  This invariant
 409exists (for example) to guarantee that if you allocate a chunk
 410which is smaller than or equal to 64 kilobytes, the extent of the
 411buffer you receive will not cross a 64K boundary.
 413To unmap and free such a DMA region, you call:
 415        dma_free_coherent(dev, size, cpu_addr, dma_handle);
 417where dev, size are the same as in the above call and cpu_addr and
 418dma_handle are the values dma_alloc_coherent() returned to you.
 419This function may not be called in interrupt context.
 421If your driver needs lots of smaller memory regions, you can write
 422custom code to subdivide pages returned by dma_alloc_coherent(),
 423or you can use the dma_pool API to do that.  A dma_pool is like
 424a kmem_cache, but it uses dma_alloc_coherent(), not __get_free_pages().
 425Also, it understands common hardware constraints for alignment,
 426like queue heads needing to be aligned on N byte boundaries.
 428Create a dma_pool like this:
 430        struct dma_pool *pool;
 432        pool = dma_pool_create(name, dev, size, align, boundary);
 434The "name" is for diagnostics (like a kmem_cache name); dev and size
 435are as above.  The device's hardware alignment requirement for this
 436type of data is "align" (which is expressed in bytes, and must be a
 437power of two).  If your device has no boundary crossing restrictions,
 438pass 0 for boundary; passing 4096 says memory allocated from this pool
 439must not cross 4KByte boundaries (but at that time it may be better to
 440use dma_alloc_coherent() directly instead).
 442Allocate memory from a DMA pool like this:
 444        cpu_addr = dma_pool_alloc(pool, flags, &dma_handle);
 446flags are GFP_KERNEL if blocking is permitted (not in_interrupt nor
 447holding SMP locks), GFP_ATOMIC otherwise.  Like dma_alloc_coherent(),
 448this returns two values, cpu_addr and dma_handle.
 450Free memory that was allocated from a dma_pool like this:
 452        dma_pool_free(pool, cpu_addr, dma_handle);
 454where pool is what you passed to dma_pool_alloc(), and cpu_addr and
 455dma_handle are the values dma_pool_alloc() returned. This function
 456may be called in interrupt context.
 458Destroy a dma_pool by calling:
 460        dma_pool_destroy(pool);
 462Make sure you've called dma_pool_free() for all memory allocated
 463from a pool before you destroy the pool. This function may not
 464be called in interrupt context.
 466                        DMA Direction
 468The interfaces described in subsequent portions of this document
 469take a DMA direction argument, which is an integer and takes on
 470one of the following values:
 477You should provide the exact DMA direction if you know it.
 479DMA_TO_DEVICE means "from main memory to the device"
 480DMA_FROM_DEVICE means "from the device to main memory"
 481It is the direction in which the data moves during the DMA
 484You are _strongly_ encouraged to specify this as precisely
 485as you possibly can.
 487If you absolutely cannot know the direction of the DMA transfer,
 488specify DMA_BIDIRECTIONAL.  It means that the DMA can go in
 489either direction.  The platform guarantees that you may legally
 490specify this, and that it will work, but this may be at the
 491cost of performance for example.
 493The value DMA_NONE is to be used for debugging.  One can
 494hold this in a data structure before you come to know the
 495precise direction, and this will help catch cases where your
 496direction tracking logic has failed to set things up properly.
 498Another advantage of specifying this value precisely (outside of
 499potential platform-specific optimizations of such) is for debugging.
 500Some platforms actually have a write permission boolean which DMA
 501mappings can be marked with, much like page protections in the user
 502program address space.  Such platforms can and do report errors in the
 503kernel logs when the DMA controller hardware detects violation of the
 504permission setting.
 506Only streaming mappings specify a direction, consistent mappings
 507implicitly have a direction attribute setting of
 510The SCSI subsystem tells you the direction to use in the
 511'sc_data_direction' member of the SCSI command your driver is
 512working on.
 514For Networking drivers, it's a rather simple affair.  For transmit
 515packets, map/unmap them with the DMA_TO_DEVICE direction
 516specifier.  For receive packets, just the opposite, map/unmap them
 517with the DMA_FROM_DEVICE direction specifier.
 519                  Using Streaming DMA mappings
 521The streaming DMA mapping routines can be called from interrupt
 522context.  There are two versions of each map/unmap, one which will
 523map/unmap a single memory region, and one which will map/unmap a
 526To map a single region, you do:
 528        struct device *dev = &my_dev->dev;
 529        dma_addr_t dma_handle;
 530        void *addr = buffer->ptr;
 531        size_t size = buffer->len;
 533        dma_handle = dma_map_single(dev, addr, size, direction);
 534        if (dma_mapping_error(dev, dma_handle)) {
 535                /*
 536                 * reduce current DMA mapping usage,
 537                 * delay and try again later or
 538                 * reset driver.
 539                 */
 540                goto map_error_handling;
 541        }
 543and to unmap it:
 545        dma_unmap_single(dev, dma_handle, size, direction);
 547You should call dma_mapping_error() as dma_map_single() could fail and return
 548error. Not all DMA implementations support the dma_mapping_error() interface.
 549However, it is a good practice to call dma_mapping_error() interface, which
 550will invoke the generic mapping error check interface. Doing so will ensure
 551that the mapping code will work correctly on all DMA implementations without
 552any dependency on the specifics of the underlying implementation. Using the
 553returned address without checking for errors could result in failures ranging
 554from panics to silent data corruption. A couple of examples of incorrect ways
 555to check for errors that make assumptions about the underlying DMA
 556implementation are as follows and these are applicable to dma_map_page() as
 559Incorrect example 1:
 560        dma_addr_t dma_handle;
 562        dma_handle = dma_map_single(dev, addr, size, direction);
 563        if ((dma_handle & 0xffff != 0) || (dma_handle >= 0x1000000)) {
 564                goto map_error;
 565        }
 567Incorrect example 2:
 568        dma_addr_t dma_handle;
 570        dma_handle = dma_map_single(dev, addr, size, direction);
 571        if (dma_handle == DMA_ERROR_CODE) {
 572                goto map_error;
 573        }
 575You should call dma_unmap_single() when the DMA activity is finished, e.g.,
 576from the interrupt which told you that the DMA transfer is done.
 578Using CPU pointers like this for single mappings has a disadvantage:
 579you cannot reference HIGHMEM memory in this way.  Thus, there is a
 580map/unmap interface pair akin to dma_{map,unmap}_single().  These
 581interfaces deal with page/offset pairs instead of CPU pointers.
 584        struct device *dev = &my_dev->dev;
 585        dma_addr_t dma_handle;
 586        struct page *page = buffer->page;
 587        unsigned long offset = buffer->offset;
 588        size_t size = buffer->len;
 590        dma_handle = dma_map_page(dev, page, offset, size, direction);
 591        if (dma_mapping_error(dev, dma_handle)) {
 592                /*
 593                 * reduce current DMA mapping usage,
 594                 * delay and try again later or
 595                 * reset driver.
 596                 */
 597                goto map_error_handling;
 598        }
 600        ...
 602        dma_unmap_page(dev, dma_handle, size, direction);
 604Here, "offset" means byte offset within the given page.
 606You should call dma_mapping_error() as dma_map_page() could fail and return
 607error as outlined under the dma_map_single() discussion.
 609You should call dma_unmap_page() when the DMA activity is finished, e.g.,
 610from the interrupt which told you that the DMA transfer is done.
 612With scatterlists, you map a region gathered from several regions by:
 614        int i, count = dma_map_sg(dev, sglist, nents, direction);
 615        struct scatterlist *sg;
 617        for_each_sg(sglist, sg, count, i) {
 618                hw_address[i] = sg_dma_address(sg);
 619                hw_len[i] = sg_dma_len(sg);
 620        }
 622where nents is the number of entries in the sglist.
 624The implementation is free to merge several consecutive sglist entries
 625into one (e.g. if DMA mapping is done with PAGE_SIZE granularity, any
 626consecutive sglist entries can be merged into one provided the first one
 627ends and the second one starts on a page boundary - in fact this is a huge
 628advantage for cards which either cannot do scatter-gather or have very
 629limited number of scatter-gather entries) and returns the actual number
 630of sg entries it mapped them to. On failure 0 is returned.
 632Then you should loop count times (note: this can be less than nents times)
 633and use sg_dma_address() and sg_dma_len() macros where you previously
 634accessed sg->address and sg->length as shown above.
 636To unmap a scatterlist, just call:
 638        dma_unmap_sg(dev, sglist, nents, direction);
 640Again, make sure DMA activity has already finished.
 642PLEASE NOTE:  The 'nents' argument to the dma_unmap_sg call must be
 643              the _same_ one you passed into the dma_map_sg call,
 644              it should _NOT_ be the 'count' value _returned_ from the
 645              dma_map_sg call.
 647Every dma_map_{single,sg}() call should have its dma_unmap_{single,sg}()
 648counterpart, because the bus address space is a shared resource and
 649you could render the machine unusable by consuming all bus addresses.
 651If you need to use the same streaming DMA region multiple times and touch
 652the data in between the DMA transfers, the buffer needs to be synced
 653properly in order for the CPU and device to see the most up-to-date and
 654correct copy of the DMA buffer.
 656So, firstly, just map it with dma_map_{single,sg}(), and after each DMA
 657transfer call either:
 659        dma_sync_single_for_cpu(dev, dma_handle, size, direction);
 663        dma_sync_sg_for_cpu(dev, sglist, nents, direction);
 665as appropriate.
 667Then, if you wish to let the device get at the DMA area again,
 668finish accessing the data with the CPU, and then before actually
 669giving the buffer to the hardware call either:
 671        dma_sync_single_for_device(dev, dma_handle, size, direction);
 675        dma_sync_sg_for_device(dev, sglist, nents, direction);
 677as appropriate.
 679After the last DMA transfer call one of the DMA unmap routines
 680dma_unmap_{single,sg}(). If you don't touch the data from the first
 681dma_map_*() call till dma_unmap_*(), then you don't have to call the
 682dma_sync_*() routines at all.
 684Here is pseudo code which shows a situation in which you would need
 685to use the dma_sync_*() interfaces.
 687        my_card_setup_receive_buffer(struct my_card *cp, char *buffer, int len)
 688        {
 689                dma_addr_t mapping;
 691                mapping = dma_map_single(cp->dev, buffer, len, DMA_FROM_DEVICE);
 692                if (dma_mapping_error(cp->dev, dma_handle)) {
 693                        /*
 694                         * reduce current DMA mapping usage,
 695                         * delay and try again later or
 696                         * reset driver.
 697                         */
 698                        goto map_error_handling;
 699                }
 701                cp->rx_buf = buffer;
 702                cp->rx_len = len;
 703                cp->rx_dma = mapping;
 705                give_rx_buf_to_card(cp);
 706        }
 708        ...
 710        my_card_interrupt_handler(int irq, void *devid, struct pt_regs *regs)
 711        {
 712                struct my_card *cp = devid;
 714                ...
 715                if (read_card_status(cp) == RX_BUF_TRANSFERRED) {
 716                        struct my_card_header *hp;
 718                        /* Examine the header to see if we wish
 719                         * to accept the data.  But synchronize
 720                         * the DMA transfer with the CPU first
 721                         * so that we see updated contents.
 722                         */
 723                        dma_sync_single_for_cpu(&cp->dev, cp->rx_dma,
 724                                                cp->rx_len,
 725                                                DMA_FROM_DEVICE);
 727                        /* Now it is safe to examine the buffer. */
 728                        hp = (struct my_card_header *) cp->rx_buf;
 729                        if (header_is_ok(hp)) {
 730                                dma_unmap_single(&cp->dev, cp->rx_dma, cp->rx_len,
 731                                                 DMA_FROM_DEVICE);
 732                                pass_to_upper_layers(cp->rx_buf);
 733                                make_and_setup_new_rx_buf(cp);
 734                        } else {
 735                                /* CPU should not write to
 736                                 * DMA_FROM_DEVICE-mapped area,
 737                                 * so dma_sync_single_for_device() is
 738                                 * not needed here. It would be required
 739                                 * for DMA_BIDIRECTIONAL mapping if
 740                                 * the memory was modified.
 741                                 */
 742                                give_rx_buf_to_card(cp);
 743                        }
 744                }
 745        }
 747Drivers converted fully to this interface should not use virt_to_bus() any
 748longer, nor should they use bus_to_virt(). Some drivers have to be changed a
 749little bit, because there is no longer an equivalent to bus_to_virt() in the
 750dynamic DMA mapping scheme - you have to always store the DMA addresses
 751returned by the dma_alloc_coherent(), dma_pool_alloc(), and dma_map_single()
 752calls (dma_map_sg() stores them in the scatterlist itself if the platform
 753supports dynamic DMA mapping in hardware) in your driver structures and/or
 754in the card registers.
 756All drivers should be using these interfaces with no exceptions.  It
 757is planned to completely remove virt_to_bus() and bus_to_virt() as
 758they are entirely deprecated.  Some ports already do not provide these
 759as it is impossible to correctly support them.
 761                        Handling Errors
 763DMA address space is limited on some architectures and an allocation
 764failure can be determined by:
 766- checking if dma_alloc_coherent() returns NULL or dma_map_sg returns 0
 768- checking the dma_addr_t returned from dma_map_single() and dma_map_page()
 769  by using dma_mapping_error():
 771        dma_addr_t dma_handle;
 773        dma_handle = dma_map_single(dev, addr, size, direction);
 774        if (dma_mapping_error(dev, dma_handle)) {
 775                /*
 776                 * reduce current DMA mapping usage,
 777                 * delay and try again later or
 778                 * reset driver.
 779                 */
 780                goto map_error_handling;
 781        }
 783- unmap pages that are already mapped, when mapping error occurs in the middle
 784  of a multiple page mapping attempt. These example are applicable to
 785  dma_map_page() as well.
 787Example 1:
 788        dma_addr_t dma_handle1;
 789        dma_addr_t dma_handle2;
 791        dma_handle1 = dma_map_single(dev, addr, size, direction);
 792        if (dma_mapping_error(dev, dma_handle1)) {
 793                /*
 794                 * reduce current DMA mapping usage,
 795                 * delay and try again later or
 796                 * reset driver.
 797                 */
 798                goto map_error_handling1;
 799        }
 800        dma_handle2 = dma_map_single(dev, addr, size, direction);
 801        if (dma_mapping_error(dev, dma_handle2)) {
 802                /*
 803                 * reduce current DMA mapping usage,
 804                 * delay and try again later or
 805                 * reset driver.
 806                 */
 807                goto map_error_handling2;
 808        }
 810        ...
 812        map_error_handling2:
 813                dma_unmap_single(dma_handle1);
 814        map_error_handling1:
 816Example 2: (if buffers are allocated in a loop, unmap all mapped buffers when
 817            mapping error is detected in the middle)
 819        dma_addr_t dma_addr;
 820        dma_addr_t array[DMA_BUFFERS];
 821        int save_index = 0;
 823        for (i = 0; i < DMA_BUFFERS; i++) {
 825                ...
 827                dma_addr = dma_map_single(dev, addr, size, direction);
 828                if (dma_mapping_error(dev, dma_addr)) {
 829                        /*
 830                         * reduce current DMA mapping usage,
 831                         * delay and try again later or
 832                         * reset driver.
 833                         */
 834                        goto map_error_handling;
 835                }
 836                array[i].dma_addr = dma_addr;
 837                save_index++;
 838        }
 840        ...
 842        map_error_handling:
 844        for (i = 0; i < save_index; i++) {
 846                ...
 848                dma_unmap_single(array[i].dma_addr);
 849        }
 851Networking drivers must call dev_kfree_skb() to free the socket buffer
 852and return NETDEV_TX_OK if the DMA mapping fails on the transmit hook
 853(ndo_start_xmit). This means that the socket buffer is just dropped in
 854the failure case.
 856SCSI drivers must return SCSI_MLQUEUE_HOST_BUSY if the DMA mapping
 857fails in the queuecommand hook. This means that the SCSI subsystem
 858passes the command to the driver again later.
 860                Optimizing Unmap State Space Consumption
 862On many platforms, dma_unmap_{single,page}() is simply a nop.
 863Therefore, keeping track of the mapping address and length is a waste
 864of space.  Instead of filling your drivers up with ifdefs and the like
 865to "work around" this (which would defeat the whole purpose of a
 866portable API) the following facilities are provided.
 868Actually, instead of describing the macros one by one, we'll
 869transform some example code.
 8711) Use DEFINE_DMA_UNMAP_{ADDR,LEN} in state saving structures.
 872   Example, before:
 874        struct ring_state {
 875                struct sk_buff *skb;
 876                dma_addr_t mapping;
 877                __u32 len;
 878        };
 880   after:
 882        struct ring_state {
 883                struct sk_buff *skb;
 884                DEFINE_DMA_UNMAP_ADDR(mapping);
 885                DEFINE_DMA_UNMAP_LEN(len);
 886        };
 8882) Use dma_unmap_{addr,len}_set() to set these values.
 889   Example, before:
 891        ringp->mapping = FOO;
 892        ringp->len = BAR;
 894   after:
 896        dma_unmap_addr_set(ringp, mapping, FOO);
 897        dma_unmap_len_set(ringp, len, BAR);
 8993) Use dma_unmap_{addr,len}() to access these values.
 900   Example, before:
 902        dma_unmap_single(dev, ringp->mapping, ringp->len,
 903                         DMA_FROM_DEVICE);
 905   after:
 907        dma_unmap_single(dev,
 908                         dma_unmap_addr(ringp, mapping),
 909                         dma_unmap_len(ringp, len),
 910                         DMA_FROM_DEVICE);
 912It really should be self-explanatory.  We treat the ADDR and LEN
 913separately, because it is possible for an implementation to only
 914need the address in order to perform the unmap operation.
 916                        Platform Issues
 918If you are just writing drivers for Linux and do not maintain
 919an architecture port for the kernel, you can safely skip down
 920to "Closing".
 9221) Struct scatterlist requirements.
 924   Don't invent the architecture specific struct scatterlist; just use
 925   <asm-generic/scatterlist.h>. You need to enable
 926   CONFIG_NEED_SG_DMA_LENGTH if the architecture supports IOMMUs
 927   (including software IOMMU).
 931   Architectures must ensure that kmalloc'ed buffer is
 932   DMA-safe. Drivers and subsystems depend on it. If an architecture
 933   isn't fully DMA-coherent (i.e. hardware doesn't ensure that data in
 934   the CPU cache is identical to data in main memory),
 935   ARCH_DMA_MINALIGN must be set so that the memory allocator
 936   makes sure that kmalloc'ed buffer doesn't share a cache line with
 937   the others. See arch/arm/include/asm/cache.h as an example.
 939   Note that ARCH_DMA_MINALIGN is about DMA memory alignment
 940   constraints. You don't need to worry about the architecture data
 941   alignment constraints (e.g. the alignment constraints about 64-bit
 942   objects).
 9443) Supporting multiple types of IOMMUs
 946   If your architecture needs to support multiple types of IOMMUs, you
 947   can use include/linux/asm-generic/dma-mapping-common.h. It's a
 948   library to support the DMA API with multiple types of IOMMUs. Lots
 949   of architectures (x86, powerpc, sh, alpha, ia64, microblaze and
 950   sparc) use it. Choose one to see how it can be used. If you need to
 951   support multiple types of IOMMUs in a single system, the example of
 952   x86 or powerpc helps.
 954                           Closing
 956This document, and the API itself, would not be in its current
 957form without the feedback and suggestions from numerous individuals.
 958We would like to specifically mention, in no particular order, the
 959following people:
 961        Russell King <>
 962        Leo Dagum <>
 963        Ralf Baechle <>
 964        Grant Grundler <>
 965        Jay Estabrook <>
 966        Thomas Sailer <>
 967        Andrea Arcangeli <>
 968        Jens Axboe <>
 969        David Mosberger-Tang <>
 970 kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.