blob: 1af0f2d5022066f1a205ad5265771e3ab112c018 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001 Dynamic DMA mapping using the generic device
2 ============================================
3
4 James E.J. Bottomley <James.Bottomley@HansenPartnership.com>
5
6This document describes the DMA API. For a more gentle introduction
7phrased in terms of the pci_ equivalents (and actual examples) see
8DMA-mapping.txt
9
10This API is split into two pieces. Part I describes the API and the
11corresponding pci_ API. Part II describes the extensions to the API
12for supporting non-consistent memory machines. Unless you know that
13your driver absolutely has to support non-consistent platforms (this
14is usually only legacy platforms) you should only use the API
15described in part I.
16
17Part I - pci_ and dma_ Equivalent API
18-------------------------------------
19
20To get the pci_ API, you must #include <linux/pci.h>
21To get the dma_ API, you must #include <linux/dma-mapping.h>
22
23
24Part Ia - Using large dma-coherent buffers
25------------------------------------------
26
27void *
28dma_alloc_coherent(struct device *dev, size_t size,
29 dma_addr_t *dma_handle, int flag)
30void *
31pci_alloc_consistent(struct pci_dev *dev, size_t size,
32 dma_addr_t *dma_handle)
33
34Consistent memory is memory for which a write by either the device or
35the processor can immediately be read by the processor or device
36without having to worry about caching effects.
37
38This routine allocates a region of <size> bytes of consistent memory.
39it also returns a <dma_handle> which may be cast to an unsigned
40integer the same width as the bus and used as the physical address
41base of the region.
42
43Returns: a pointer to the allocated region (in the processor's virtual
44address space) or NULL if the allocation failed.
45
46Note: consistent memory can be expensive on some platforms, and the
47minimum allocation length may be as big as a page, so you should
48consolidate your requests for consistent memory as much as possible.
49The simplest way to do that is to use the dma_pool calls (see below).
50
51The flag parameter (dma_alloc_coherent only) allows the caller to
52specify the GFP_ flags (see kmalloc) for the allocation (the
53implementation may chose to ignore flags that affect the location of
54the returned memory, like GFP_DMA). For pci_alloc_consistent, you
55must assume GFP_ATOMIC behaviour.
56
57void
58dma_free_coherent(struct device *dev, size_t size, void *cpu_addr
59 dma_addr_t dma_handle)
60void
61pci_free_consistent(struct pci_dev *dev, size_t size, void *cpu_addr
62 dma_addr_t dma_handle)
63
64Free the region of consistent memory you previously allocated. dev,
65size and dma_handle must all be the same as those passed into the
66consistent allocate. cpu_addr must be the virtual address returned by
67the consistent allocate
68
69
70Part Ib - Using small dma-coherent buffers
71------------------------------------------
72
73To get this part of the dma_ API, you must #include <linux/dmapool.h>
74
75Many drivers need lots of small dma-coherent memory regions for DMA
76descriptors or I/O buffers. Rather than allocating in units of a page
77or more using dma_alloc_coherent(), you can use DMA pools. These work
78much like a kmem_cache_t, except that they use the dma-coherent allocator
79not __get_free_pages(). Also, they understand common hardware constraints
80for alignment, like queue heads needing to be aligned on N byte boundaries.
81
82
83 struct dma_pool *
84 dma_pool_create(const char *name, struct device *dev,
85 size_t size, size_t align, size_t alloc);
86
87 struct pci_pool *
88 pci_pool_create(const char *name, struct pci_device *dev,
89 size_t size, size_t align, size_t alloc);
90
91The pool create() routines initialize a pool of dma-coherent buffers
92for use with a given device. It must be called in a context which
93can sleep.
94
95The "name" is for diagnostics (like a kmem_cache_t name); dev and size
96are like what you'd pass to dma_alloc_coherent(). The device's hardware
97alignment requirement for this type of data is "align" (which is expressed
98in bytes, and must be a power of two). If your device has no boundary
99crossing restrictions, pass 0 for alloc; passing 4096 says memory allocated
100from this pool must not cross 4KByte boundaries.
101
102
103 void *dma_pool_alloc(struct dma_pool *pool, int gfp_flags,
104 dma_addr_t *dma_handle);
105
106 void *pci_pool_alloc(struct pci_pool *pool, int gfp_flags,
107 dma_addr_t *dma_handle);
108
109This allocates memory from the pool; the returned memory will meet the size
110and alignment requirements specified at creation time. Pass GFP_ATOMIC to
111prevent blocking, or if it's permitted (not in_interrupt, not holding SMP locks)
112pass GFP_KERNEL to allow blocking. Like dma_alloc_coherent(), this returns
113two values: an address usable by the cpu, and the dma address usable by the
114pool's device.
115
116
117 void dma_pool_free(struct dma_pool *pool, void *vaddr,
118 dma_addr_t addr);
119
120 void pci_pool_free(struct pci_pool *pool, void *vaddr,
121 dma_addr_t addr);
122
123This puts memory back into the pool. The pool is what was passed to
Tobias Klauserd533f672005-09-10 00:26:46 -0700124the pool allocation routine; the cpu and dma addresses are what
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125were returned when that routine allocated the memory being freed.
126
127
128 void dma_pool_destroy(struct dma_pool *pool);
129
130 void pci_pool_destroy(struct pci_pool *pool);
131
132The pool destroy() routines free the resources of the pool. They must be
133called in a context which can sleep. Make sure you've freed all allocated
134memory back to the pool before you destroy it.
135
136
137Part Ic - DMA addressing limitations
138------------------------------------
139
140int
141dma_supported(struct device *dev, u64 mask)
142int
143pci_dma_supported(struct device *dev, u64 mask)
144
145Checks to see if the device can support DMA to the memory described by
146mask.
147
148Returns: 1 if it can and 0 if it can't.
149
150Notes: This routine merely tests to see if the mask is possible. It
151won't change the current mask settings. It is more intended as an
152internal API for use by the platform than an external API for use by
153driver writers.
154
155int
156dma_set_mask(struct device *dev, u64 mask)
157int
158pci_set_dma_mask(struct pci_device *dev, u64 mask)
159
160Checks to see if the mask is possible and updates the device
161parameters if it is.
162
163Returns: 0 if successful and a negative error if not.
164
165u64
166dma_get_required_mask(struct device *dev)
167
168After setting the mask with dma_set_mask(), this API returns the
169actual mask (within that already set) that the platform actually
170requires to operate efficiently. Usually this means the returned mask
171is the minimum required to cover all of memory. Examining the
172required mask gives drivers with variable descriptor sizes the
173opportunity to use smaller descriptors as necessary.
174
175Requesting the required mask does not alter the current mask. If you
176wish to take advantage of it, you should issue another dma_set_mask()
177call to lower the mask again.
178
179
180Part Id - Streaming DMA mappings
181--------------------------------
182
183dma_addr_t
184dma_map_single(struct device *dev, void *cpu_addr, size_t size,
185 enum dma_data_direction direction)
186dma_addr_t
187pci_map_single(struct device *dev, void *cpu_addr, size_t size,
188 int direction)
189
190Maps a piece of processor virtual memory so it can be accessed by the
191device and returns the physical handle of the memory.
192
193The direction for both api's may be converted freely by casting.
194However the dma_ API uses a strongly typed enumerator for its
195direction:
196
197DMA_NONE = PCI_DMA_NONE no direction (used for
198 debugging)
199DMA_TO_DEVICE = PCI_DMA_TODEVICE data is going from the
200 memory to the device
201DMA_FROM_DEVICE = PCI_DMA_FROMDEVICE data is coming from
202 the device to the
203 memory
204DMA_BIDIRECTIONAL = PCI_DMA_BIDIRECTIONAL direction isn't known
205
206Notes: Not all memory regions in a machine can be mapped by this
207API. Further, regions that appear to be physically contiguous in
208kernel virtual space may not be contiguous as physical memory. Since
209this API does not provide any scatter/gather capability, it will fail
210if the user tries to map a non physically contiguous piece of memory.
211For this reason, it is recommended that memory mapped by this API be
212obtained only from sources which guarantee to be physically contiguous
213(like kmalloc).
214
215Further, the physical address of the memory must be within the
216dma_mask of the device (the dma_mask represents a bit mask of the
217addressable region for the device. i.e. if the physical address of
218the memory anded with the dma_mask is still equal to the physical
219address, then the device can perform DMA to the memory). In order to
220ensure that the memory allocated by kmalloc is within the dma_mask,
221the driver may specify various platform dependent flags to restrict
222the physical memory range of the allocation (e.g. on x86, GFP_DMA
223guarantees to be within the first 16Mb of available physical memory,
224as required by ISA devices).
225
226Note also that the above constraints on physical contiguity and
227dma_mask may not apply if the platform has an IOMMU (a device which
228supplies a physical to virtual mapping between the I/O memory bus and
229the device). However, to be portable, device driver writers may *not*
230assume that such an IOMMU exists.
231
232Warnings: Memory coherency operates at a granularity called the cache
233line width. In order for memory mapped by this API to operate
234correctly, the mapped region must begin exactly on a cache line
235boundary and end exactly on one (to prevent two separately mapped
236regions from sharing a single cache line). Since the cache line size
237may not be known at compile time, the API will not enforce this
238requirement. Therefore, it is recommended that driver writers who
239don't take special care to determine the cache line size at run time
240only map virtual regions that begin and end on page boundaries (which
241are guaranteed also to be cache line boundaries).
242
243DMA_TO_DEVICE synchronisation must be done after the last modification
244of the memory region by the software and before it is handed off to
245the driver. Once this primitive is used. Memory covered by this
246primitive should be treated as read only by the device. If the device
247may write to it at any point, it should be DMA_BIDIRECTIONAL (see
248below).
249
250DMA_FROM_DEVICE synchronisation must be done before the driver
251accesses data that may be changed by the device. This memory should
252be treated as read only by the driver. If the driver needs to write
253to it at any point, it should be DMA_BIDIRECTIONAL (see below).
254
255DMA_BIDIRECTIONAL requires special handling: it means that the driver
256isn't sure if the memory was modified before being handed off to the
257device and also isn't sure if the device will also modify it. Thus,
258you must always sync bidirectional memory twice: once before the
259memory is handed off to the device (to make sure all memory changes
260are flushed from the processor) and once before the data may be
261accessed after being used by the device (to make sure any processor
262cache lines are updated with data that the device may have changed.
263
264void
265dma_unmap_single(struct device *dev, dma_addr_t dma_addr, size_t size,
266 enum dma_data_direction direction)
267void
268pci_unmap_single(struct pci_dev *hwdev, dma_addr_t dma_addr,
269 size_t size, int direction)
270
271Unmaps the region previously mapped. All the parameters passed in
272must be identical to those passed in (and returned) by the mapping
273API.
274
275dma_addr_t
276dma_map_page(struct device *dev, struct page *page,
277 unsigned long offset, size_t size,
278 enum dma_data_direction direction)
279dma_addr_t
280pci_map_page(struct pci_dev *hwdev, struct page *page,
281 unsigned long offset, size_t size, int direction)
282void
283dma_unmap_page(struct device *dev, dma_addr_t dma_address, size_t size,
284 enum dma_data_direction direction)
285void
286pci_unmap_page(struct pci_dev *hwdev, dma_addr_t dma_address,
287 size_t size, int direction)
288
289API for mapping and unmapping for pages. All the notes and warnings
290for the other mapping APIs apply here. Also, although the <offset>
291and <size> parameters are provided to do partial page mapping, it is
292recommended that you never use these unless you really know what the
293cache width is.
294
295int
296dma_mapping_error(dma_addr_t dma_addr)
297
298int
299pci_dma_mapping_error(dma_addr_t dma_addr)
300
301In some circumstances dma_map_single and dma_map_page will fail to create
302a mapping. A driver can check for these errors by testing the returned
303dma address with dma_mapping_error(). A non zero return value means the mapping
304could not be created and the driver should take appropriate action (eg
305reduce current DMA mapping usage or delay and try again later).
306
307int
308dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
309 enum dma_data_direction direction)
310int
311pci_map_sg(struct pci_dev *hwdev, struct scatterlist *sg,
312 int nents, int direction)
313
314Maps a scatter gather list from the block layer.
315
316Returns: the number of physical segments mapped (this may be shorted
317than <nents> passed in if the block layer determines that some
318elements of the scatter/gather list are physically adjacent and thus
319may be mapped with a single entry).
320
321Please note that the sg cannot be mapped again if it has been mapped once.
322The mapping process is allowed to destroy information in the sg.
323
324As with the other mapping interfaces, dma_map_sg can fail. When it
325does, 0 is returned and a driver must take appropriate action. It is
326critical that the driver do something, in the case of a block driver
327aborting the request or even oopsing is better than doing nothing and
328corrupting the filesystem.
329
330void
331dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nhwentries,
332 enum dma_data_direction direction)
333void
334pci_unmap_sg(struct pci_dev *hwdev, struct scatterlist *sg,
335 int nents, int direction)
336
337unmap the previously mapped scatter/gather list. All the parameters
338must be the same as those and passed in to the scatter/gather mapping
339API.
340
341Note: <nents> must be the number you passed in, *not* the number of
342physical entries returned.
343
344void
345dma_sync_single(struct device *dev, dma_addr_t dma_handle, size_t size,
346 enum dma_data_direction direction)
347void
348pci_dma_sync_single(struct pci_dev *hwdev, dma_addr_t dma_handle,
349 size_t size, int direction)
350void
351dma_sync_sg(struct device *dev, struct scatterlist *sg, int nelems,
352 enum dma_data_direction direction)
353void
354pci_dma_sync_sg(struct pci_dev *hwdev, struct scatterlist *sg,
355 int nelems, int direction)
356
357synchronise a single contiguous or scatter/gather mapping. All the
358parameters must be the same as those passed into the single mapping
359API.
360
361Notes: You must do this:
362
363- Before reading values that have been written by DMA from the device
364 (use the DMA_FROM_DEVICE direction)
365- After writing values that will be written to the device using DMA
366 (use the DMA_TO_DEVICE) direction
367- before *and* after handing memory to the device if the memory is
368 DMA_BIDIRECTIONAL
369
370See also dma_map_single().
371
372
373Part II - Advanced dma_ usage
374-----------------------------
375
376Warning: These pieces of the DMA API have no PCI equivalent. They
377should also not be used in the majority of cases, since they cater for
378unlikely corner cases that don't belong in usual drivers.
379
380If you don't understand how cache line coherency works between a
381processor and an I/O device, you should not be using this part of the
382API at all.
383
384void *
385dma_alloc_noncoherent(struct device *dev, size_t size,
386 dma_addr_t *dma_handle, int flag)
387
388Identical to dma_alloc_coherent() except that the platform will
389choose to return either consistent or non-consistent memory as it sees
390fit. By using this API, you are guaranteeing to the platform that you
391have all the correct and necessary sync points for this memory in the
392driver should it choose to return non-consistent memory.
393
394Note: where the platform can return consistent memory, it will
395guarantee that the sync points become nops.
396
397Warning: Handling non-consistent memory is a real pain. You should
398only ever use this API if you positively know your driver will be
399required to work on one of the rare (usually non-PCI) architectures
400that simply cannot make consistent memory.
401
402void
403dma_free_noncoherent(struct device *dev, size_t size, void *cpu_addr,
404 dma_addr_t dma_handle)
405
406free memory allocated by the nonconsistent API. All parameters must
407be identical to those passed in (and returned by
408dma_alloc_noncoherent()).
409
410int
411dma_is_consistent(dma_addr_t dma_handle)
412
413returns true if the memory pointed to by the dma_handle is actually
414consistent.
415
416int
417dma_get_cache_alignment(void)
418
419returns the processor cache alignment. This is the absolute minimum
420alignment *and* width that you must observe when either mapping
421memory or doing partial flushes.
422
423Notes: This API may return a number *larger* than the actual cache
424line, but it will guarantee that one or more cache lines fit exactly
425into the width returned by this call. It will also always be a power
426of two for easy alignment
427
428void
429dma_sync_single_range(struct device *dev, dma_addr_t dma_handle,
430 unsigned long offset, size_t size,
431 enum dma_data_direction direction)
432
433does a partial sync. starting at offset and continuing for size. You
434must be careful to observe the cache alignment and width when doing
435anything like this. You must also be extra careful about accessing
436memory you intend to sync partially.
437
438void
439dma_cache_sync(void *vaddr, size_t size,
440 enum dma_data_direction direction)
441
442Do a partial sync of memory that was allocated by
443dma_alloc_noncoherent(), starting at virtual address vaddr and
444continuing on for size. Again, you *must* observe the cache line
445boundaries when doing this.
446
447int
448dma_declare_coherent_memory(struct device *dev, dma_addr_t bus_addr,
449 dma_addr_t device_addr, size_t size, int
450 flags)
451
452
453Declare region of memory to be handed out by dma_alloc_coherent when
454it's asked for coherent memory for this device.
455
456bus_addr is the physical address to which the memory is currently
457assigned in the bus responding region (this will be used by the
458platform to perform the mapping)
459
460device_addr is the physical address the device needs to be programmed
461with actually to address this memory (this will be handed out as the
462dma_addr_t in dma_alloc_coherent())
463
464size is the size of the area (must be multiples of PAGE_SIZE).
465
466flags can be or'd together and are
467
468DMA_MEMORY_MAP - request that the memory returned from
469dma_alloc_coherent() be directly writeable.
470
471DMA_MEMORY_IO - request that the memory returned from
472dma_alloc_coherent() be addressable using read/write/memcpy_toio etc.
473
474One or both of these flags must be present
475
476DMA_MEMORY_INCLUDES_CHILDREN - make the declared memory be allocated by
477dma_alloc_coherent of any child devices of this one (for memory residing
478on a bridge).
479
480DMA_MEMORY_EXCLUSIVE - only allocate memory from the declared regions.
481Do not allow dma_alloc_coherent() to fall back to system memory when
482it's out of memory in the declared region.
483
484The return value will be either DMA_MEMORY_MAP or DMA_MEMORY_IO and
485must correspond to a passed in flag (i.e. no returning DMA_MEMORY_IO
486if only DMA_MEMORY_MAP were passed in) for success or zero for
487failure.
488
489Note, for DMA_MEMORY_IO returns, all subsequent memory returned by
490dma_alloc_coherent() may no longer be accessed directly, but instead
491must be accessed using the correct bus functions. If your driver
492isn't prepared to handle this contingency, it should not specify
493DMA_MEMORY_IO in the input flags.
494
495As a simplification for the platforms, only *one* such region of
496memory may be declared per device.
497
498For reasons of efficiency, most platforms choose to track the declared
499region only at the granularity of a page. For smaller allocations,
500you should use the dma_pool() API.
501
502void
503dma_release_declared_memory(struct device *dev)
504
505Remove the memory region previously declared from the system. This
506API performs *no* in-use checking for this region and will return
507unconditionally having removed all the required structures. It is the
508drivers job to ensure that no parts of this memory region are
509currently in use.
510
511void *
512dma_mark_declared_memory_occupied(struct device *dev,
513 dma_addr_t device_addr, size_t size)
514
515This is used to occupy specific regions of the declared space
516(dma_alloc_coherent() will hand out the first free region it finds).
517
518device_addr is the *device* address of the region requested
519
520size is the size (and should be a page sized multiple).
521
522The return value will be either a pointer to the processor virtual
523address of the memory, or an error (via PTR_ERR()) if any part of the
524region is occupied.
525
526