dma.rst (6156B)
1USB DMA 2~~~~~~~ 3 4In Linux 2.5 kernels (and later), USB device drivers have additional control 5over how DMA may be used to perform I/O operations. The APIs are detailed 6in the kernel usb programming guide (kerneldoc, from the source code). 7 8API overview 9============ 10 11The big picture is that USB drivers can continue to ignore most DMA issues, 12though they still must provide DMA-ready buffers (see 13Documentation/core-api/dma-api-howto.rst). That's how they've worked through 14the 2.4 (and earlier) kernels, or they can now be DMA-aware. 15 16DMA-aware usb drivers: 17 18- New calls enable DMA-aware drivers, letting them allocate dma buffers and 19 manage dma mappings for existing dma-ready buffers (see below). 20 21- URBs have an additional "transfer_dma" field, as well as a transfer_flags 22 bit saying if it's valid. (Control requests also have "setup_dma", but 23 drivers must not use it.) 24 25- "usbcore" will map this DMA address, if a DMA-aware driver didn't do 26 it first and set ``URB_NO_TRANSFER_DMA_MAP``. HCDs 27 don't manage dma mappings for URBs. 28 29- There's a new "generic DMA API", parts of which are usable by USB device 30 drivers. Never use dma_set_mask() on any USB interface or device; that 31 would potentially break all devices sharing that bus. 32 33Eliminating copies 34================== 35 36It's good to avoid making CPUs copy data needlessly. The costs can add up, 37and effects like cache-trashing can impose subtle penalties. 38 39- If you're doing lots of small data transfers from the same buffer all 40 the time, that can really burn up resources on systems which use an 41 IOMMU to manage the DMA mappings. It can cost MUCH more to set up and 42 tear down the IOMMU mappings with each request than perform the I/O! 43 44 For those specific cases, USB has primitives to allocate less expensive 45 memory. They work like kmalloc and kfree versions that give you the right 46 kind of addresses to store in urb->transfer_buffer and urb->transfer_dma. 47 You'd also set ``URB_NO_TRANSFER_DMA_MAP`` in urb->transfer_flags:: 48 49 void *usb_alloc_coherent (struct usb_device *dev, size_t size, 50 int mem_flags, dma_addr_t *dma); 51 52 void usb_free_coherent (struct usb_device *dev, size_t size, 53 void *addr, dma_addr_t dma); 54 55 Most drivers should **NOT** be using these primitives; they don't need 56 to use this type of memory ("dma-coherent"), and memory returned from 57 :c:func:`kmalloc` will work just fine. 58 59 The memory buffer returned is "dma-coherent"; sometimes you might need to 60 force a consistent memory access ordering by using memory barriers. It's 61 not using a streaming DMA mapping, so it's good for small transfers on 62 systems where the I/O would otherwise thrash an IOMMU mapping. (See 63 Documentation/core-api/dma-api-howto.rst for definitions of "coherent" and 64 "streaming" DMA mappings.) 65 66 Asking for 1/Nth of a page (as well as asking for N pages) is reasonably 67 space-efficient. 68 69 On most systems the memory returned will be uncached, because the 70 semantics of dma-coherent memory require either bypassing CPU caches 71 or using cache hardware with bus-snooping support. While x86 hardware 72 has such bus-snooping, many other systems use software to flush cache 73 lines to prevent DMA conflicts. 74 75- Devices on some EHCI controllers could handle DMA to/from high memory. 76 77 Unfortunately, the current Linux DMA infrastructure doesn't have a sane 78 way to expose these capabilities ... and in any case, HIGHMEM is mostly a 79 design wart specific to x86_32. So your best bet is to ensure you never 80 pass a highmem buffer into a USB driver. That's easy; it's the default 81 behavior. Just don't override it; e.g. with ``NETIF_F_HIGHDMA``. 82 83 This may force your callers to do some bounce buffering, copying from 84 high memory to "normal" DMA memory. If you can come up with a good way 85 to fix this issue (for x86_32 machines with over 1 GByte of memory), 86 feel free to submit patches. 87 88Working with existing buffers 89============================= 90 91Existing buffers aren't usable for DMA without first being mapped into the 92DMA address space of the device. However, most buffers passed to your 93driver can safely be used with such DMA mapping. (See the first section 94of Documentation/core-api/dma-api-howto.rst, titled "What memory is DMA-able?") 95 96- When you're using scatterlists, you can map everything at once. On some 97 systems, this kicks in an IOMMU and turns the scatterlists into single 98 DMA transactions:: 99 100 int usb_buffer_map_sg (struct usb_device *dev, unsigned pipe, 101 struct scatterlist *sg, int nents); 102 103 void usb_buffer_dmasync_sg (struct usb_device *dev, unsigned pipe, 104 struct scatterlist *sg, int n_hw_ents); 105 106 void usb_buffer_unmap_sg (struct usb_device *dev, unsigned pipe, 107 struct scatterlist *sg, int n_hw_ents); 108 109 It's probably easier to use the new ``usb_sg_*()`` calls, which do the DMA 110 mapping and apply other tweaks to make scatterlist i/o be fast. 111 112- Some drivers may prefer to work with the model that they're mapping large 113 buffers, synchronizing their safe re-use. (If there's no re-use, then let 114 usbcore do the map/unmap.) Large periodic transfers make good examples 115 here, since it's cheaper to just synchronize the buffer than to unmap it 116 each time an urb completes and then re-map it on during resubmission. 117 118 These calls all work with initialized urbs: ``urb->dev``, ``urb->pipe``, 119 ``urb->transfer_buffer``, and ``urb->transfer_buffer_length`` must all be 120 valid when these calls are used (``urb->setup_packet`` must be valid too 121 if urb is a control request):: 122 123 struct urb *usb_buffer_map (struct urb *urb); 124 125 void usb_buffer_dmasync (struct urb *urb); 126 127 void usb_buffer_unmap (struct urb *urb); 128 129 The calls manage ``urb->transfer_dma`` for you, and set 130 ``URB_NO_TRANSFER_DMA_MAP`` so that usbcore won't map or unmap the buffer. 131 They cannot be used for setup_packet buffers in control requests. 132 133Note that several of those interfaces are currently commented out, since 134they don't have current users. See the source code. Other than the dmasync 135calls (where the underlying DMA primitives have changed), most of them can 136easily be commented back in if you want to use them.