4 This document describes the semantics of the DMA attributes that are
5 defined in linux/dma-attrs.h.
10 DMA_ATTR_WRITE_BARRIER is a (write) barrier attribute for DMA. DMA
11 to a memory region with the DMA_ATTR_WRITE_BARRIER attribute forces
12 all pending DMA writes to complete, and thus provides a mechanism to
13 strictly order DMA from a device across all intervening busses and
14 bridges. This barrier is not specific to a particular type of
15 interconnect, it applies to the system as a whole, and so its
16 implementation must account for the idiosyncrasies of the system all
17 the way from the DMA device to memory.
19 As an example of a situation where DMA_ATTR_WRITE_BARRIER would be
20 useful, suppose that a device does a DMA write to indicate that data is
21 ready and available in memory. The DMA of the "completion indication"
22 could race with data DMA. Mapping the memory used for completion
23 indications with DMA_ATTR_WRITE_BARRIER would prevent the race.
25 DMA_ATTR_WEAK_ORDERING
26 ----------------------
28 DMA_ATTR_WEAK_ORDERING specifies that reads and writes to the mapping
29 may be weakly ordered, that is that reads and writes may pass each other.
31 Since it is optional for platforms to implement DMA_ATTR_WEAK_ORDERING,
32 those that do not will simply ignore the attribute and exhibit default
35 DMA_ATTR_WRITE_COMBINE
36 ----------------------
38 DMA_ATTR_WRITE_COMBINE specifies that writes to the mapping may be
39 buffered to improve performance.
41 Since it is optional for platforms to implement DMA_ATTR_WRITE_COMBINE,
42 those that do not will simply ignore the attribute and exhibit default
45 DMA_ATTR_NON_CONSISTENT
46 -----------------------
48 DMA_ATTR_NON_CONSISTENT lets the platform to choose to return either
49 consistent or non-consistent memory as it sees fit. By using this API,
50 you are guaranteeing to the platform that you have all the correct and
51 necessary sync points for this memory in the driver.
53 DMA_ATTR_NO_KERNEL_MAPPING
54 --------------------------
56 DMA_ATTR_NO_KERNEL_MAPPING lets the platform to avoid creating a kernel
57 virtual mapping for the allocated buffer. On some architectures creating
58 such mapping is non-trivial task and consumes very limited resources
59 (like kernel virtual address space or dma consistent address space).
60 Buffers allocated with this attribute can be only passed to user space
61 by calling dma_mmap_attrs(). By using this API, you are guaranteeing
62 that you won't dereference the pointer returned by dma_alloc_attr(). You
63 can treat it as a cookie that must be passed to dma_mmap_attrs() and
64 dma_free_attrs(). Make sure that both of these also get this attribute
67 Since it is optional for platforms to implement
68 DMA_ATTR_NO_KERNEL_MAPPING, those that do not will simply ignore the
69 attribute and exhibit default behavior.
71 DMA_ATTR_SKIP_CPU_SYNC
72 ----------------------
74 By default dma_map_{single,page,sg} functions family transfer a given
75 buffer from CPU domain to device domain. Some advanced use cases might
76 require sharing a buffer between more than one device. This requires
77 having a mapping created separately for each device and is usually
78 performed by calling dma_map_{single,page,sg} function more than once
79 for the given buffer with device pointer to each device taking part in
80 the buffer sharing. The first call transfers a buffer from 'CPU' domain
81 to 'device' domain, what synchronizes CPU caches for the given region
82 (usually it means that the cache has been flushed or invalidated
83 depending on the dma direction). However, next calls to
84 dma_map_{single,page,sg}() for other devices will perform exactly the
85 same synchronization operation on the CPU cache. CPU cache synchronization
86 might be a time consuming operation, especially if the buffers are
87 large, so it is highly recommended to avoid it if possible.
88 DMA_ATTR_SKIP_CPU_SYNC allows platform code to skip synchronization of
89 the CPU cache for the given buffer assuming that it has been already
90 transferred to 'device' domain. This attribute can be also used for
91 dma_unmap_{single,page,sg} functions family to force buffer to stay in
92 device domain after releasing a mapping for it. Use this attribute with
95 DMA_ATTR_FORCE_CONTIGUOUS
96 -------------------------
98 By default DMA-mapping subsystem is allowed to assemble the buffer
99 allocated by dma_alloc_attrs() function from individual pages if it can
100 be mapped as contiguous chunk into device dma address space. By
101 specifying this attribute the allocated buffer is forced to be contiguous
102 also in physical memory.
104 DMA_ATTR_ALLOC_SINGLE_PAGES
105 ---------------------------
107 This is a hint to the DMA-mapping subsystem that it's probably not worth
108 the time to try to allocate memory to in a way that gives better TLB
109 efficiency (AKA it's not worth trying to build the mapping out of larger
110 pages). You might want to specify this if:
111 - You know that the accesses to this memory won't thrash the TLB.
112 You might know that the accesses are likely to be sequential or
113 that they aren't sequential but it's unlikely you'll ping-pong
114 between many addresses that are likely to be in different physical
116 - You know that the penalty of TLB misses while accessing the
117 memory will be small enough to be inconsequential. If you are
118 doing a heavy operation like decryption or decompression this
120 - You know that the DMA mapping is fairly transitory. If you expect
121 the mapping to have a short lifetime then it may be worth it to
122 optimize allocation (avoid coming up with large pages) instead of
123 getting the slight performance win of larger pages.
124 Setting this hint doesn't guarantee that you won't get huge pages, but it
125 means that we won't try quite as hard to get them.
127 NOTE: At the moment DMA_ATTR_ALLOC_SINGLE_PAGES is only implemented on ARM,
128 though ARM64 patches will likely be posted soon.