Root/Documentation/io-mapping.txt

1The io_mapping functions in linux/io-mapping.h provide an abstraction for
2efficiently mapping small regions of an I/O device to the CPU. The initial
3usage is to support the large graphics aperture on 32-bit processors where
4ioremap_wc cannot be used to statically map the entire aperture to the CPU
5as it would consume too much of the kernel address space.
6
7A mapping object is created during driver initialization using
8
9    struct io_mapping *io_mapping_create_wc(unsigned long base,
10                        unsigned long size)
11
12        'base' is the bus address of the region to be made
13        mappable, while 'size' indicates how large a mapping region to
14        enable. Both are in bytes.
15
16        This _wc variant provides a mapping which may only be used
17        with the io_mapping_map_atomic_wc or io_mapping_map_wc.
18
19With this mapping object, individual pages can be mapped either atomically
20or not, depending on the necessary scheduling environment. Of course, atomic
21maps are more efficient:
22
23    void *io_mapping_map_atomic_wc(struct io_mapping *mapping,
24                       unsigned long offset)
25
26        'offset' is the offset within the defined mapping region.
27        Accessing addresses beyond the region specified in the
28        creation function yields undefined results. Using an offset
29        which is not page aligned yields an undefined result. The
30        return value points to a single page in CPU address space.
31
32        This _wc variant returns a write-combining map to the
33        page and may only be used with mappings created by
34        io_mapping_create_wc
35
36        Note that the task may not sleep while holding this page
37        mapped.
38
39    void io_mapping_unmap_atomic(void *vaddr)
40
41        'vaddr' must be the the value returned by the last
42        io_mapping_map_atomic_wc call. This unmaps the specified
43        page and allows the task to sleep once again.
44
45If you need to sleep while holding the lock, you can use the non-atomic
46variant, although they may be significantly slower.
47
48    void *io_mapping_map_wc(struct io_mapping *mapping,
49                unsigned long offset)
50
51        This works like io_mapping_map_atomic_wc except it allows
52        the task to sleep while holding the page mapped.
53
54    void io_mapping_unmap(void *vaddr)
55
56        This works like io_mapping_unmap_atomic, except it is used
57        for pages mapped with io_mapping_map_wc.
58
59At driver close time, the io_mapping object must be freed:
60
61    void io_mapping_free(struct io_mapping *mapping)
62
63Current Implementation:
64
65The initial implementation of these functions uses existing mapping
66mechanisms and so provides only an abstraction layer and no new
67functionality.
68
69On 64-bit processors, io_mapping_create_wc calls ioremap_wc for the whole
70range, creating a permanent kernel-visible mapping to the resource. The
71map_atomic and map functions add the requested offset to the base of the
72virtual address returned by ioremap_wc.
73
74On 32-bit processors with HIGHMEM defined, io_mapping_map_atomic_wc uses
75kmap_atomic_pfn to map the specified page in an atomic fashion;
76kmap_atomic_pfn isn't really supposed to be used with device pages, but it
77provides an efficient mapping for this usage.
78
79On 32-bit processors without HIGHMEM defined, io_mapping_map_atomic_wc and
80io_mapping_map_wc both use ioremap_wc, a terribly inefficient function which
81performs an IPI to inform all processors about the new mapping. This results
82in a significant performance penalty.
83

Archive Download this file



interactive