1In Linux 2.5 kernels (and later), USB device drivers have additional control
   2over how DMA may be used to perform I/O operations.  The APIs are detailed
   3in the kernel usb programming guide (kerneldoc, from the source code).
   8The big picture is that USB drivers can continue to ignore most DMA issues,
   9though they still must provide DMA-ready buffers (see
  10Documentation/DMA-API-HOWTO.txt).  That's how they've worked through
  11the 2.4 (and earlier) kernels.
  13OR:  they can now be DMA-aware.
  15- New calls enable DMA-aware drivers, letting them allocate dma buffers and
  16  manage dma mappings for existing dma-ready buffers (see below).
  18- URBs have an additional "transfer_dma" field, as well as a transfer_flags
  19  bit saying if it's valid.  (Control requests also have "setup_dma", but
  20  drivers must not use it.)
  22- "usbcore" will map this DMA address, if a DMA-aware driver didn't do
  23  it first and set URB_NO_TRANSFER_DMA_MAP.  HCDs
  24  don't manage dma mappings for URBs.
  26- There's a new "generic DMA API", parts of which are usable by USB device
  27  drivers.  Never use dma_set_mask() on any USB interface or device; that
  28  would potentially break all devices sharing that bus.
  33It's good to avoid making CPUs copy data needlessly.  The costs can add up,
  34and effects like cache-trashing can impose subtle penalties.
  36- If you're doing lots of small data transfers from the same buffer all
  37  the time, that can really burn up resources on systems which use an
  38  IOMMU to manage the DMA mappings.  It can cost MUCH more to set up and
  39  tear down the IOMMU mappings with each request than perform the I/O!
  41  For those specific cases, USB has primitives to allocate less expensive
  42  memory.  They work like kmalloc and kfree versions that give you the right
  43  kind of addresses to store in urb->transfer_buffer and urb->transfer_dma.
  44  You'd also set URB_NO_TRANSFER_DMA_MAP in urb->transfer_flags:
  46        void *usb_alloc_coherent (struct usb_device *dev, size_t size,
  47                int mem_flags, dma_addr_t *dma);
  49        void usb_free_coherent (struct usb_device *dev, size_t size,
  50                void *addr, dma_addr_t dma);
  52  Most drivers should *NOT* be using these primitives; they don't need
  53  to use this type of memory ("dma-coherent"), and memory returned from
  54  kmalloc() will work just fine.
  56  The memory buffer returned is "dma-coherent"; sometimes you might need to
  57  force a consistent memory access ordering by using memory barriers.  It's
  58  not using a streaming DMA mapping, so it's good for small transfers on
  59  systems where the I/O would otherwise thrash an IOMMU mapping.  (See
  60  Documentation/DMA-API-HOWTO.txt for definitions of "coherent" and
  61  "streaming" DMA mappings.)
  63  Asking for 1/Nth of a page (as well as asking for N pages) is reasonably
  64  space-efficient.
  66  On most systems the memory returned will be uncached, because the
  67  semantics of dma-coherent memory require either bypassing CPU caches
  68  or using cache hardware with bus-snooping support.  While x86 hardware
  69  has such bus-snooping, many other systems use software to flush cache
  70  lines to prevent DMA conflicts.
  72- Devices on some EHCI controllers could handle DMA to/from high memory.
  74  Unfortunately, the current Linux DMA infrastructure doesn't have a sane
  75  way to expose these capabilities ... and in any case, HIGHMEM is mostly a
  76  design wart specific to x86_32.  So your best bet is to ensure you never
  77  pass a highmem buffer into a USB driver.  That's easy; it's the default
  78  behavior.  Just don't override it; e.g. with NETIF_F_HIGHDMA.
  80  This may force your callers to do some bounce buffering, copying from
  81  high memory to "normal" DMA memory.  If you can come up with a good way
  82  to fix this issue (for x86_32 machines with over 1 GByte of memory),
  83  feel free to submit patches.
  88Existing buffers aren't usable for DMA without first being mapped into the
  89DMA address space of the device.  However, most buffers passed to your
  90driver can safely be used with such DMA mapping.  (See the first section
  91of Documentation/DMA-API-HOWTO.txt, titled "What memory is DMA-able?")
  93- When you're using scatterlists, you can map everything at once.  On some
  94  systems, this kicks in an IOMMU and turns the scatterlists into single
  95  DMA transactions:
  97        int usb_buffer_map_sg (struct usb_device *dev, unsigned pipe,
  98                struct scatterlist *sg, int nents);
 100        void usb_buffer_dmasync_sg (struct usb_device *dev, unsigned pipe,
 101                struct scatterlist *sg, int n_hw_ents);
 103        void usb_buffer_unmap_sg (struct usb_device *dev, unsigned pipe,
 104                struct scatterlist *sg, int n_hw_ents);
 106  It's probably easier to use the new usb_sg_*() calls, which do the DMA
 107  mapping and apply other tweaks to make scatterlist i/o be fast.
 109- Some drivers may prefer to work with the model that they're mapping large
 110  buffers, synchronizing their safe re-use.  (If there's no re-use, then let
 111  usbcore do the map/unmap.)  Large periodic transfers make good examples
 112  here, since it's cheaper to just synchronize the buffer than to unmap it
 113  each time an urb completes and then re-map it on during resubmission.
 115  These calls all work with initialized urbs:  urb->dev, urb->pipe,
 116  urb->transfer_buffer, and urb->transfer_buffer_length must all be
 117  valid when these calls are used (urb->setup_packet must be valid too
 118  if urb is a control request):
 120        struct urb *usb_buffer_map (struct urb *urb);
 122        void usb_buffer_dmasync (struct urb *urb);
 124        void usb_buffer_unmap (struct urb *urb);
 126  The calls manage urb->transfer_dma for you, and set URB_NO_TRANSFER_DMA_MAP
 127  so that usbcore won't map or unmap the buffer.  They cannot be used for
 128  setup_packet buffers in control requests.
 130Note that several of those interfaces are currently commented out, since
 131they don't have current users.  See the source code.  Other than the dmasync
 132calls (where the underlying DMA primitives have changed), most of them can
 133easily be commented back in if you want to use them.