Loading...
1 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
7of the API (and actual examples) see
8Documentation/DMA-API-HOWTO.txt.
9
10This API is split into two pieces. Part I describes the API. Part II
11describes the extensions to the API for supporting non-consistent
12memory machines. Unless you know that your driver absolutely has to
13support non-consistent platforms (this is usually only legacy
14platforms) you should only use the API described in part I.
15
16Part I - dma_ API
17-------------------------------------
18
19To get the dma_ API, you must #include <linux/dma-mapping.h>
20
21
22Part Ia - Using large dma-coherent buffers
23------------------------------------------
24
25void *
26dma_alloc_coherent(struct device *dev, size_t size,
27 dma_addr_t *dma_handle, gfp_t flag)
28
29Consistent memory is memory for which a write by either the device or
30the processor can immediately be read by the processor or device
31without having to worry about caching effects. (You may however need
32to make sure to flush the processor's write buffers before telling
33devices to read that memory.)
34
35This routine allocates a region of <size> bytes of consistent memory.
36It also returns a <dma_handle> which may be cast to an unsigned
37integer the same width as the bus and used as the physical address
38base of the region.
39
40Returns: a pointer to the allocated region (in the processor's virtual
41address space) or NULL if the allocation failed.
42
43Note: consistent memory can be expensive on some platforms, and the
44minimum allocation length may be as big as a page, so you should
45consolidate your requests for consistent memory as much as possible.
46The simplest way to do that is to use the dma_pool calls (see below).
47
48The flag parameter (dma_alloc_coherent only) allows the caller to
49specify the GFP_ flags (see kmalloc) for the allocation (the
50implementation may choose to ignore flags that affect the location of
51the returned memory, like GFP_DMA).
52
53void
54dma_free_coherent(struct device *dev, size_t size, void *cpu_addr,
55 dma_addr_t dma_handle)
56
57Free the region of consistent memory you previously allocated. dev,
58size and dma_handle must all be the same as those passed into the
59consistent allocate. cpu_addr must be the virtual address returned by
60the consistent allocate.
61
62Note that unlike their sibling allocation calls, these routines
63may only be called with IRQs enabled.
64
65
66Part Ib - Using small dma-coherent buffers
67------------------------------------------
68
69To get this part of the dma_ API, you must #include <linux/dmapool.h>
70
71Many drivers need lots of small dma-coherent memory regions for DMA
72descriptors or I/O buffers. Rather than allocating in units of a page
73or more using dma_alloc_coherent(), you can use DMA pools. These work
74much like a struct kmem_cache, except that they use the dma-coherent allocator,
75not __get_free_pages(). Also, they understand common hardware constraints
76for alignment, like queue heads needing to be aligned on N-byte boundaries.
77
78
79 struct dma_pool *
80 dma_pool_create(const char *name, struct device *dev,
81 size_t size, size_t align, size_t alloc);
82
83The pool create() routines initialize a pool of dma-coherent buffers
84for use with a given device. It must be called in a context which
85can sleep.
86
87The "name" is for diagnostics (like a struct kmem_cache name); dev and size
88are like what you'd pass to dma_alloc_coherent(). The device's hardware
89alignment requirement for this type of data is "align" (which is expressed
90in bytes, and must be a power of two). If your device has no boundary
91crossing restrictions, pass 0 for alloc; passing 4096 says memory allocated
92from this pool must not cross 4KByte boundaries.
93
94
95 void *dma_pool_alloc(struct dma_pool *pool, gfp_t gfp_flags,
96 dma_addr_t *dma_handle);
97
98This allocates memory from the pool; the returned memory will meet the size
99and alignment requirements specified at creation time. Pass GFP_ATOMIC to
100prevent blocking, or if it's permitted (not in_interrupt, not holding SMP locks),
101pass GFP_KERNEL to allow blocking. Like dma_alloc_coherent(), this returns
102two values: an address usable by the cpu, and the dma address usable by the
103pool's device.
104
105
106 void dma_pool_free(struct dma_pool *pool, void *vaddr,
107 dma_addr_t addr);
108
109This puts memory back into the pool. The pool is what was passed to
110the pool allocation routine; the cpu (vaddr) and dma addresses are what
111were returned when that routine allocated the memory being freed.
112
113
114 void dma_pool_destroy(struct dma_pool *pool);
115
116The pool destroy() routines free the resources of the pool. They must be
117called in a context which can sleep. Make sure you've freed all allocated
118memory back to the pool before you destroy it.
119
120
121Part Ic - DMA addressing limitations
122------------------------------------
123
124int
125dma_supported(struct device *dev, u64 mask)
126
127Checks to see if the device can support DMA to the memory described by
128mask.
129
130Returns: 1 if it can and 0 if it can't.
131
132Notes: This routine merely tests to see if the mask is possible. It
133won't change the current mask settings. It is more intended as an
134internal API for use by the platform than an external API for use by
135driver writers.
136
137int
138dma_set_mask(struct device *dev, u64 mask)
139
140Checks to see if the mask is possible and updates the device
141parameters if it is.
142
143Returns: 0 if successful and a negative error if not.
144
145int
146dma_set_coherent_mask(struct device *dev, u64 mask)
147
148Checks to see if the mask is possible and updates the device
149parameters if it is.
150
151Returns: 0 if successful and a negative error if not.
152
153u64
154dma_get_required_mask(struct device *dev)
155
156This API returns the mask that the platform requires to
157operate efficiently. Usually this means the returned mask
158is the minimum required to cover all of memory. Examining the
159required mask gives drivers with variable descriptor sizes the
160opportunity to use smaller descriptors as necessary.
161
162Requesting the required mask does not alter the current mask. If you
163wish to take advantage of it, you should issue a dma_set_mask()
164call to set the mask to the value returned.
165
166
167Part Id - Streaming DMA mappings
168--------------------------------
169
170dma_addr_t
171dma_map_single(struct device *dev, void *cpu_addr, size_t size,
172 enum dma_data_direction direction)
173
174Maps a piece of processor virtual memory so it can be accessed by the
175device and returns the physical handle of the memory.
176
177The direction for both api's may be converted freely by casting.
178However the dma_ API uses a strongly typed enumerator for its
179direction:
180
181DMA_NONE no direction (used for debugging)
182DMA_TO_DEVICE data is going from the memory to the device
183DMA_FROM_DEVICE data is coming from the device to the memory
184DMA_BIDIRECTIONAL direction isn't known
185
186Notes: Not all memory regions in a machine can be mapped by this
187API. Further, regions that appear to be physically contiguous in
188kernel virtual space may not be contiguous as physical memory. Since
189this API does not provide any scatter/gather capability, it will fail
190if the user tries to map a non-physically contiguous piece of memory.
191For this reason, it is recommended that memory mapped by this API be
192obtained only from sources which guarantee it to be physically contiguous
193(like kmalloc).
194
195Further, the physical address of the memory must be within the
196dma_mask of the device (the dma_mask represents a bit mask of the
197addressable region for the device. I.e., if the physical address of
198the memory anded with the dma_mask is still equal to the physical
199address, then the device can perform DMA to the memory). In order to
200ensure that the memory allocated by kmalloc is within the dma_mask,
201the driver may specify various platform-dependent flags to restrict
202the physical memory range of the allocation (e.g. on x86, GFP_DMA
203guarantees to be within the first 16Mb of available physical memory,
204as required by ISA devices).
205
206Note also that the above constraints on physical contiguity and
207dma_mask may not apply if the platform has an IOMMU (a device which
208supplies a physical to virtual mapping between the I/O memory bus and
209the device). However, to be portable, device driver writers may *not*
210assume that such an IOMMU exists.
211
212Warnings: Memory coherency operates at a granularity called the cache
213line width. In order for memory mapped by this API to operate
214correctly, the mapped region must begin exactly on a cache line
215boundary and end exactly on one (to prevent two separately mapped
216regions from sharing a single cache line). Since the cache line size
217may not be known at compile time, the API will not enforce this
218requirement. Therefore, it is recommended that driver writers who
219don't take special care to determine the cache line size at run time
220only map virtual regions that begin and end on page boundaries (which
221are guaranteed also to be cache line boundaries).
222
223DMA_TO_DEVICE synchronisation must be done after the last modification
224of the memory region by the software and before it is handed off to
225the driver. Once this primitive is used, memory covered by this
226primitive should be treated as read-only by the device. If the device
227may write to it at any point, it should be DMA_BIDIRECTIONAL (see
228below).
229
230DMA_FROM_DEVICE synchronisation must be done before the driver
231accesses data that may be changed by the device. This memory should
232be treated as read-only by the driver. If the driver needs to write
233to it at any point, it should be DMA_BIDIRECTIONAL (see below).
234
235DMA_BIDIRECTIONAL requires special handling: it means that the driver
236isn't sure if the memory was modified before being handed off to the
237device and also isn't sure if the device will also modify it. Thus,
238you must always sync bidirectional memory twice: once before the
239memory is handed off to the device (to make sure all memory changes
240are flushed from the processor) and once before the data may be
241accessed after being used by the device (to make sure any processor
242cache lines are updated with data that the device may have changed).
243
244void
245dma_unmap_single(struct device *dev, dma_addr_t dma_addr, size_t size,
246 enum dma_data_direction direction)
247
248Unmaps the region previously mapped. All the parameters passed in
249must be identical to those passed in (and returned) by the mapping
250API.
251
252dma_addr_t
253dma_map_page(struct device *dev, struct page *page,
254 unsigned long offset, size_t size,
255 enum dma_data_direction direction)
256void
257dma_unmap_page(struct device *dev, dma_addr_t dma_address, size_t size,
258 enum dma_data_direction direction)
259
260API for mapping and unmapping for pages. All the notes and warnings
261for the other mapping APIs apply here. Also, although the <offset>
262and <size> parameters are provided to do partial page mapping, it is
263recommended that you never use these unless you really know what the
264cache width is.
265
266int
267dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
268
269In some circumstances dma_map_single and dma_map_page will fail to create
270a mapping. A driver can check for these errors by testing the returned
271dma address with dma_mapping_error(). A non-zero return value means the mapping
272could not be created and the driver should take appropriate action (e.g.
273reduce current DMA mapping usage or delay and try again later).
274
275 int
276 dma_map_sg(struct device *dev, struct scatterlist *sg,
277 int nents, enum dma_data_direction direction)
278
279Returns: the number of physical segments mapped (this may be shorter
280than <nents> passed in if some elements of the scatter/gather list are
281physically or virtually adjacent and an IOMMU maps them with a single
282entry).
283
284Please note that the sg cannot be mapped again if it has been mapped once.
285The mapping process is allowed to destroy information in the sg.
286
287As with the other mapping interfaces, dma_map_sg can fail. When it
288does, 0 is returned and a driver must take appropriate action. It is
289critical that the driver do something, in the case of a block driver
290aborting the request or even oopsing is better than doing nothing and
291corrupting the filesystem.
292
293With scatterlists, you use the resulting mapping like this:
294
295 int i, count = dma_map_sg(dev, sglist, nents, direction);
296 struct scatterlist *sg;
297
298 for_each_sg(sglist, sg, count, i) {
299 hw_address[i] = sg_dma_address(sg);
300 hw_len[i] = sg_dma_len(sg);
301 }
302
303where nents is the number of entries in the sglist.
304
305The implementation is free to merge several consecutive sglist entries
306into one (e.g. with an IOMMU, or if several pages just happen to be
307physically contiguous) and returns the actual number of sg entries it
308mapped them to. On failure 0, is returned.
309
310Then you should loop count times (note: this can be less than nents times)
311and use sg_dma_address() and sg_dma_len() macros where you previously
312accessed sg->address and sg->length as shown above.
313
314 void
315 dma_unmap_sg(struct device *dev, struct scatterlist *sg,
316 int nhwentries, enum dma_data_direction direction)
317
318Unmap the previously mapped scatter/gather list. All the parameters
319must be the same as those and passed in to the scatter/gather mapping
320API.
321
322Note: <nents> must be the number you passed in, *not* the number of
323physical entries returned.
324
325void
326dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle, size_t size,
327 enum dma_data_direction direction)
328void
329dma_sync_single_for_device(struct device *dev, dma_addr_t dma_handle, size_t size,
330 enum dma_data_direction direction)
331void
332dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg, int nelems,
333 enum dma_data_direction direction)
334void
335dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg, int nelems,
336 enum dma_data_direction direction)
337
338Synchronise a single contiguous or scatter/gather mapping for the cpu
339and device. With the sync_sg API, all the parameters must be the same
340as those passed into the single mapping API. With the sync_single API,
341you can use dma_handle and size parameters that aren't identical to
342those passed into the single mapping API to do a partial sync.
343
344Notes: You must do this:
345
346- Before reading values that have been written by DMA from the device
347 (use the DMA_FROM_DEVICE direction)
348- After writing values that will be written to the device using DMA
349 (use the DMA_TO_DEVICE) direction
350- before *and* after handing memory to the device if the memory is
351 DMA_BIDIRECTIONAL
352
353See also dma_map_single().
354
355dma_addr_t
356dma_map_single_attrs(struct device *dev, void *cpu_addr, size_t size,
357 enum dma_data_direction dir,
358 struct dma_attrs *attrs)
359
360void
361dma_unmap_single_attrs(struct device *dev, dma_addr_t dma_addr,
362 size_t size, enum dma_data_direction dir,
363 struct dma_attrs *attrs)
364
365int
366dma_map_sg_attrs(struct device *dev, struct scatterlist *sgl,
367 int nents, enum dma_data_direction dir,
368 struct dma_attrs *attrs)
369
370void
371dma_unmap_sg_attrs(struct device *dev, struct scatterlist *sgl,
372 int nents, enum dma_data_direction dir,
373 struct dma_attrs *attrs)
374
375The four functions above are just like the counterpart functions
376without the _attrs suffixes, except that they pass an optional
377struct dma_attrs*.
378
379struct dma_attrs encapsulates a set of "dma attributes". For the
380definition of struct dma_attrs see linux/dma-attrs.h.
381
382The interpretation of dma attributes is architecture-specific, and
383each attribute should be documented in Documentation/DMA-attributes.txt.
384
385If struct dma_attrs* is NULL, the semantics of each of these
386functions is identical to those of the corresponding function
387without the _attrs suffix. As a result dma_map_single_attrs()
388can generally replace dma_map_single(), etc.
389
390As an example of the use of the *_attrs functions, here's how
391you could pass an attribute DMA_ATTR_FOO when mapping memory
392for DMA:
393
394#include <linux/dma-attrs.h>
395/* DMA_ATTR_FOO should be defined in linux/dma-attrs.h and
396 * documented in Documentation/DMA-attributes.txt */
397...
398
399 DEFINE_DMA_ATTRS(attrs);
400 dma_set_attr(DMA_ATTR_FOO, &attrs);
401 ....
402 n = dma_map_sg_attrs(dev, sg, nents, DMA_TO_DEVICE, &attr);
403 ....
404
405Architectures that care about DMA_ATTR_FOO would check for its
406presence in their implementations of the mapping and unmapping
407routines, e.g.:
408
409void whizco_dma_map_sg_attrs(struct device *dev, dma_addr_t dma_addr,
410 size_t size, enum dma_data_direction dir,
411 struct dma_attrs *attrs)
412{
413 ....
414 int foo = dma_get_attr(DMA_ATTR_FOO, attrs);
415 ....
416 if (foo)
417 /* twizzle the frobnozzle */
418 ....
419
420
421Part II - Advanced dma_ usage
422-----------------------------
423
424Warning: These pieces of the DMA API should not be used in the
425majority of cases, since they cater for unlikely corner cases that
426don't belong in usual drivers.
427
428If you don't understand how cache line coherency works between a
429processor and an I/O device, you should not be using this part of the
430API at all.
431
432void *
433dma_alloc_noncoherent(struct device *dev, size_t size,
434 dma_addr_t *dma_handle, gfp_t flag)
435
436Identical to dma_alloc_coherent() except that the platform will
437choose to return either consistent or non-consistent memory as it sees
438fit. By using this API, you are guaranteeing to the platform that you
439have all the correct and necessary sync points for this memory in the
440driver should it choose to return non-consistent memory.
441
442Note: where the platform can return consistent memory, it will
443guarantee that the sync points become nops.
444
445Warning: Handling non-consistent memory is a real pain. You should
446only ever use this API if you positively know your driver will be
447required to work on one of the rare (usually non-PCI) architectures
448that simply cannot make consistent memory.
449
450void
451dma_free_noncoherent(struct device *dev, size_t size, void *cpu_addr,
452 dma_addr_t dma_handle)
453
454Free memory allocated by the nonconsistent API. All parameters must
455be identical to those passed in (and returned by
456dma_alloc_noncoherent()).
457
458int
459dma_get_cache_alignment(void)
460
461Returns the processor cache alignment. This is the absolute minimum
462alignment *and* width that you must observe when either mapping
463memory or doing partial flushes.
464
465Notes: This API may return a number *larger* than the actual cache
466line, but it will guarantee that one or more cache lines fit exactly
467into the width returned by this call. It will also always be a power
468of two for easy alignment.
469
470void
471dma_cache_sync(struct device *dev, void *vaddr, size_t size,
472 enum dma_data_direction direction)
473
474Do a partial sync of memory that was allocated by
475dma_alloc_noncoherent(), starting at virtual address vaddr and
476continuing on for size. Again, you *must* observe the cache line
477boundaries when doing this.
478
479int
480dma_declare_coherent_memory(struct device *dev, dma_addr_t bus_addr,
481 dma_addr_t device_addr, size_t size, int
482 flags)
483
484Declare region of memory to be handed out by dma_alloc_coherent when
485it's asked for coherent memory for this device.
486
487bus_addr is the physical address to which the memory is currently
488assigned in the bus responding region (this will be used by the
489platform to perform the mapping).
490
491device_addr is the physical address the device needs to be programmed
492with actually to address this memory (this will be handed out as the
493dma_addr_t in dma_alloc_coherent()).
494
495size is the size of the area (must be multiples of PAGE_SIZE).
496
497flags can be or'd together and are:
498
499DMA_MEMORY_MAP - request that the memory returned from
500dma_alloc_coherent() be directly writable.
501
502DMA_MEMORY_IO - request that the memory returned from
503dma_alloc_coherent() be addressable using read/write/memcpy_toio etc.
504
505One or both of these flags must be present.
506
507DMA_MEMORY_INCLUDES_CHILDREN - make the declared memory be allocated by
508dma_alloc_coherent of any child devices of this one (for memory residing
509on a bridge).
510
511DMA_MEMORY_EXCLUSIVE - only allocate memory from the declared regions.
512Do not allow dma_alloc_coherent() to fall back to system memory when
513it's out of memory in the declared region.
514
515The return value will be either DMA_MEMORY_MAP or DMA_MEMORY_IO and
516must correspond to a passed in flag (i.e. no returning DMA_MEMORY_IO
517if only DMA_MEMORY_MAP were passed in) for success or zero for
518failure.
519
520Note, for DMA_MEMORY_IO returns, all subsequent memory returned by
521dma_alloc_coherent() may no longer be accessed directly, but instead
522must be accessed using the correct bus functions. If your driver
523isn't prepared to handle this contingency, it should not specify
524DMA_MEMORY_IO in the input flags.
525
526As a simplification for the platforms, only *one* such region of
527memory may be declared per device.
528
529For reasons of efficiency, most platforms choose to track the declared
530region only at the granularity of a page. For smaller allocations,
531you should use the dma_pool() API.
532
533void
534dma_release_declared_memory(struct device *dev)
535
536Remove the memory region previously declared from the system. This
537API performs *no* in-use checking for this region and will return
538unconditionally having removed all the required structures. It is the
539driver's job to ensure that no parts of this memory region are
540currently in use.
541
542void *
543dma_mark_declared_memory_occupied(struct device *dev,
544 dma_addr_t device_addr, size_t size)
545
546This is used to occupy specific regions of the declared space
547(dma_alloc_coherent() will hand out the first free region it finds).
548
549device_addr is the *device* address of the region requested.
550
551size is the size (and should be a page-sized multiple).
552
553The return value will be either a pointer to the processor virtual
554address of the memory, or an error (via PTR_ERR()) if any part of the
555region is occupied.
556
557Part III - Debug drivers use of the DMA-API
558-------------------------------------------
559
560The DMA-API as described above as some constraints. DMA addresses must be
561released with the corresponding function with the same size for example. With
562the advent of hardware IOMMUs it becomes more and more important that drivers
563do not violate those constraints. In the worst case such a violation can
564result in data corruption up to destroyed filesystems.
565
566To debug drivers and find bugs in the usage of the DMA-API checking code can
567be compiled into the kernel which will tell the developer about those
568violations. If your architecture supports it you can select the "Enable
569debugging of DMA-API usage" option in your kernel configuration. Enabling this
570option has a performance impact. Do not enable it in production kernels.
571
572If you boot the resulting kernel will contain code which does some bookkeeping
573about what DMA memory was allocated for which device. If this code detects an
574error it prints a warning message with some details into your kernel log. An
575example warning message may look like this:
576
577------------[ cut here ]------------
578WARNING: at /data2/repos/linux-2.6-iommu/lib/dma-debug.c:448
579 check_unmap+0x203/0x490()
580Hardware name:
581forcedeth 0000:00:08.0: DMA-API: device driver frees DMA memory with wrong
582 function [device address=0x00000000640444be] [size=66 bytes] [mapped as
583single] [unmapped as page]
584Modules linked in: nfsd exportfs bridge stp llc r8169
585Pid: 0, comm: swapper Tainted: G W 2.6.28-dmatest-09289-g8bb99c0 #1
586Call Trace:
587 <IRQ> [<ffffffff80240b22>] warn_slowpath+0xf2/0x130
588 [<ffffffff80647b70>] _spin_unlock+0x10/0x30
589 [<ffffffff80537e75>] usb_hcd_link_urb_to_ep+0x75/0xc0
590 [<ffffffff80647c22>] _spin_unlock_irqrestore+0x12/0x40
591 [<ffffffff8055347f>] ohci_urb_enqueue+0x19f/0x7c0
592 [<ffffffff80252f96>] queue_work+0x56/0x60
593 [<ffffffff80237e10>] enqueue_task_fair+0x20/0x50
594 [<ffffffff80539279>] usb_hcd_submit_urb+0x379/0xbc0
595 [<ffffffff803b78c3>] cpumask_next_and+0x23/0x40
596 [<ffffffff80235177>] find_busiest_group+0x207/0x8a0
597 [<ffffffff8064784f>] _spin_lock_irqsave+0x1f/0x50
598 [<ffffffff803c7ea3>] check_unmap+0x203/0x490
599 [<ffffffff803c8259>] debug_dma_unmap_page+0x49/0x50
600 [<ffffffff80485f26>] nv_tx_done_optimized+0xc6/0x2c0
601 [<ffffffff80486c13>] nv_nic_irq_optimized+0x73/0x2b0
602 [<ffffffff8026df84>] handle_IRQ_event+0x34/0x70
603 [<ffffffff8026ffe9>] handle_edge_irq+0xc9/0x150
604 [<ffffffff8020e3ab>] do_IRQ+0xcb/0x1c0
605 [<ffffffff8020c093>] ret_from_intr+0x0/0xa
606 <EOI> <4>---[ end trace f6435a98e2a38c0e ]---
607
608The driver developer can find the driver and the device including a stacktrace
609of the DMA-API call which caused this warning.
610
611Per default only the first error will result in a warning message. All other
612errors will only silently counted. This limitation exist to prevent the code
613from flooding your kernel log. To support debugging a device driver this can
614be disabled via debugfs. See the debugfs interface documentation below for
615details.
616
617The debugfs directory for the DMA-API debugging code is called dma-api/. In
618this directory the following files can currently be found:
619
620 dma-api/all_errors This file contains a numeric value. If this
621 value is not equal to zero the debugging code
622 will print a warning for every error it finds
623 into the kernel log. Be careful with this
624 option, as it can easily flood your logs.
625
626 dma-api/disabled This read-only file contains the character 'Y'
627 if the debugging code is disabled. This can
628 happen when it runs out of memory or if it was
629 disabled at boot time
630
631 dma-api/error_count This file is read-only and shows the total
632 numbers of errors found.
633
634 dma-api/num_errors The number in this file shows how many
635 warnings will be printed to the kernel log
636 before it stops. This number is initialized to
637 one at system boot and be set by writing into
638 this file
639
640 dma-api/min_free_entries
641 This read-only file can be read to get the
642 minimum number of free dma_debug_entries the
643 allocator has ever seen. If this value goes
644 down to zero the code will disable itself
645 because it is not longer reliable.
646
647 dma-api/num_free_entries
648 The current number of free dma_debug_entries
649 in the allocator.
650
651 dma-api/driver-filter
652 You can write a name of a driver into this file
653 to limit the debug output to requests from that
654 particular driver. Write an empty string to
655 that file to disable the filter and see
656 all errors again.
657
658If you have this code compiled into your kernel it will be enabled by default.
659If you want to boot without the bookkeeping anyway you can provide
660'dma_debug=off' as a boot parameter. This will disable DMA-API debugging.
661Notice that you can not enable it again at runtime. You have to reboot to do
662so.
663
664If you want to see debug messages only for a special device driver you can
665specify the dma_debug_driver=<drivername> parameter. This will enable the
666driver filter at boot time. The debug code will only print errors for that
667driver afterwards. This filter can be disabled or changed later using debugfs.
668
669When the code disables itself at runtime this is most likely because it ran
670out of dma_debug_entries. These entries are preallocated at boot. The number
671of preallocated entries is defined per architecture. If it is too low for you
672boot with 'dma_debug_entries=<your_desired_number>' to overwrite the
673architectural default.
1 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
7of the API (and actual examples), see Documentation/DMA-API-HOWTO.txt.
8
9This API is split into two pieces. Part I describes the basic API.
10Part II describes extensions for supporting non-consistent memory
11machines. Unless you know that your driver absolutely has to support
12non-consistent platforms (this is usually only legacy platforms) you
13should only use the API described in part I.
14
15Part I - dma_ API
16-------------------------------------
17
18To get the dma_ API, you must #include <linux/dma-mapping.h>. This
19provides dma_addr_t and the interfaces described below.
20
21A dma_addr_t can hold any valid DMA address for the platform. It can be
22given to a device to use as a DMA source or target. A CPU cannot reference
23a dma_addr_t directly because there may be translation between its physical
24address space and the DMA address space.
25
26Part Ia - Using large DMA-coherent buffers
27------------------------------------------
28
29void *
30dma_alloc_coherent(struct device *dev, size_t size,
31 dma_addr_t *dma_handle, gfp_t flag)
32
33Consistent memory is memory for which a write by either the device or
34the processor can immediately be read by the processor or device
35without having to worry about caching effects. (You may however need
36to make sure to flush the processor's write buffers before telling
37devices to read that memory.)
38
39This routine allocates a region of <size> bytes of consistent memory.
40
41It returns a pointer to the allocated region (in the processor's virtual
42address space) or NULL if the allocation failed.
43
44It also returns a <dma_handle> which may be cast to an unsigned integer the
45same width as the bus and given to the device as the DMA address base of
46the region.
47
48Note: consistent memory can be expensive on some platforms, and the
49minimum allocation length may be as big as a page, so you should
50consolidate your requests for consistent memory as much as possible.
51The simplest way to do that is to use the dma_pool calls (see below).
52
53The flag parameter (dma_alloc_coherent() only) allows the caller to
54specify the GFP_ flags (see kmalloc()) for the allocation (the
55implementation may choose to ignore flags that affect the location of
56the returned memory, like GFP_DMA).
57
58void *
59dma_zalloc_coherent(struct device *dev, size_t size,
60 dma_addr_t *dma_handle, gfp_t flag)
61
62Wraps dma_alloc_coherent() and also zeroes the returned memory if the
63allocation attempt succeeded.
64
65void
66dma_free_coherent(struct device *dev, size_t size, void *cpu_addr,
67 dma_addr_t dma_handle)
68
69Free a region of consistent memory you previously allocated. dev,
70size and dma_handle must all be the same as those passed into
71dma_alloc_coherent(). cpu_addr must be the virtual address returned by
72the dma_alloc_coherent().
73
74Note that unlike their sibling allocation calls, these routines
75may only be called with IRQs enabled.
76
77
78Part Ib - Using small DMA-coherent buffers
79------------------------------------------
80
81To get this part of the dma_ API, you must #include <linux/dmapool.h>
82
83Many drivers need lots of small DMA-coherent memory regions for DMA
84descriptors or I/O buffers. Rather than allocating in units of a page
85or more using dma_alloc_coherent(), you can use DMA pools. These work
86much like a struct kmem_cache, except that they use the DMA-coherent allocator,
87not __get_free_pages(). Also, they understand common hardware constraints
88for alignment, like queue heads needing to be aligned on N-byte boundaries.
89
90
91 struct dma_pool *
92 dma_pool_create(const char *name, struct device *dev,
93 size_t size, size_t align, size_t alloc);
94
95dma_pool_create() initializes a pool of DMA-coherent buffers
96for use with a given device. It must be called in a context which
97can sleep.
98
99The "name" is for diagnostics (like a struct kmem_cache name); dev and size
100are like what you'd pass to dma_alloc_coherent(). The device's hardware
101alignment requirement for this type of data is "align" (which is expressed
102in bytes, and must be a power of two). If your device has no boundary
103crossing restrictions, pass 0 for alloc; passing 4096 says memory allocated
104from this pool must not cross 4KByte boundaries.
105
106
107 void *dma_pool_zalloc(struct dma_pool *pool, gfp_t mem_flags,
108 dma_addr_t *handle)
109
110Wraps dma_pool_alloc() and also zeroes the returned memory if the
111allocation attempt succeeded.
112
113
114 void *dma_pool_alloc(struct dma_pool *pool, gfp_t gfp_flags,
115 dma_addr_t *dma_handle);
116
117This allocates memory from the pool; the returned memory will meet the
118size and alignment requirements specified at creation time. Pass
119GFP_ATOMIC to prevent blocking, or if it's permitted (not
120in_interrupt, not holding SMP locks), pass GFP_KERNEL to allow
121blocking. Like dma_alloc_coherent(), this returns two values: an
122address usable by the CPU, and the DMA address usable by the pool's
123device.
124
125
126 void dma_pool_free(struct dma_pool *pool, void *vaddr,
127 dma_addr_t addr);
128
129This puts memory back into the pool. The pool is what was passed to
130dma_pool_alloc(); the CPU (vaddr) and DMA addresses are what
131were returned when that routine allocated the memory being freed.
132
133
134 void dma_pool_destroy(struct dma_pool *pool);
135
136dma_pool_destroy() frees the resources of the pool. It must be
137called in a context which can sleep. Make sure you've freed all allocated
138memory back to the pool before you destroy it.
139
140
141Part Ic - DMA addressing limitations
142------------------------------------
143
144int
145dma_set_mask_and_coherent(struct device *dev, u64 mask)
146
147Checks to see if the mask is possible and updates the device
148streaming and coherent DMA mask parameters if it is.
149
150Returns: 0 if successful and a negative error if not.
151
152int
153dma_set_mask(struct device *dev, u64 mask)
154
155Checks to see if the mask is possible and updates the device
156parameters if it is.
157
158Returns: 0 if successful and a negative error if not.
159
160int
161dma_set_coherent_mask(struct device *dev, u64 mask)
162
163Checks to see if the mask is possible and updates the device
164parameters if it is.
165
166Returns: 0 if successful and a negative error if not.
167
168u64
169dma_get_required_mask(struct device *dev)
170
171This API returns the mask that the platform requires to
172operate efficiently. Usually this means the returned mask
173is the minimum required to cover all of memory. Examining the
174required mask gives drivers with variable descriptor sizes the
175opportunity to use smaller descriptors as necessary.
176
177Requesting the required mask does not alter the current mask. If you
178wish to take advantage of it, you should issue a dma_set_mask()
179call to set the mask to the value returned.
180
181
182Part Id - Streaming DMA mappings
183--------------------------------
184
185dma_addr_t
186dma_map_single(struct device *dev, void *cpu_addr, size_t size,
187 enum dma_data_direction direction)
188
189Maps a piece of processor virtual memory so it can be accessed by the
190device and returns the DMA address of the memory.
191
192The direction for both APIs may be converted freely by casting.
193However the dma_ API uses a strongly typed enumerator for its
194direction:
195
196DMA_NONE no direction (used for debugging)
197DMA_TO_DEVICE data is going from the memory to the device
198DMA_FROM_DEVICE data is coming from the device to the memory
199DMA_BIDIRECTIONAL direction isn't known
200
201Notes: Not all memory regions in a machine can be mapped by this API.
202Further, contiguous kernel virtual space may not be contiguous as
203physical memory. Since this API does not provide any scatter/gather
204capability, it will fail if the user tries to map a non-physically
205contiguous piece of memory. For this reason, memory to be mapped by
206this API should be obtained from sources which guarantee it to be
207physically contiguous (like kmalloc).
208
209Further, the DMA address of the memory must be within the
210dma_mask of the device (the dma_mask is a bit mask of the
211addressable region for the device, i.e., if the DMA address of
212the memory ANDed with the dma_mask is still equal to the DMA
213address, then the device can perform DMA to the memory). To
214ensure that the memory allocated by kmalloc is within the dma_mask,
215the driver may specify various platform-dependent flags to restrict
216the DMA address range of the allocation (e.g., on x86, GFP_DMA
217guarantees to be within the first 16MB of available DMA addresses,
218as required by ISA devices).
219
220Note also that the above constraints on physical contiguity and
221dma_mask may not apply if the platform has an IOMMU (a device which
222maps an I/O DMA address to a physical memory address). However, to be
223portable, device driver writers may *not* assume that such an IOMMU
224exists.
225
226Warnings: Memory coherency operates at a granularity called the cache
227line width. In order for memory mapped by this API to operate
228correctly, the mapped region must begin exactly on a cache line
229boundary and end exactly on one (to prevent two separately mapped
230regions from sharing a single cache line). Since the cache line size
231may not be known at compile time, the API will not enforce this
232requirement. Therefore, it is recommended that driver writers who
233don't take special care to determine the cache line size at run time
234only map virtual regions that begin and end on page boundaries (which
235are guaranteed also to be cache line boundaries).
236
237DMA_TO_DEVICE synchronisation must be done after the last modification
238of the memory region by the software and before it is handed off to
239the device. Once this primitive is used, memory covered by this
240primitive should be treated as read-only by the device. If the device
241may write to it at any point, it should be DMA_BIDIRECTIONAL (see
242below).
243
244DMA_FROM_DEVICE synchronisation must be done before the driver
245accesses data that may be changed by the device. This memory should
246be treated as read-only by the driver. If the driver needs to write
247to it at any point, it should be DMA_BIDIRECTIONAL (see below).
248
249DMA_BIDIRECTIONAL requires special handling: it means that the driver
250isn't sure if the memory was modified before being handed off to the
251device and also isn't sure if the device will also modify it. Thus,
252you must always sync bidirectional memory twice: once before the
253memory is handed off to the device (to make sure all memory changes
254are flushed from the processor) and once before the data may be
255accessed after being used by the device (to make sure any processor
256cache lines are updated with data that the device may have changed).
257
258void
259dma_unmap_single(struct device *dev, dma_addr_t dma_addr, size_t size,
260 enum dma_data_direction direction)
261
262Unmaps the region previously mapped. All the parameters passed in
263must be identical to those passed in (and returned) by the mapping
264API.
265
266dma_addr_t
267dma_map_page(struct device *dev, struct page *page,
268 unsigned long offset, size_t size,
269 enum dma_data_direction direction)
270void
271dma_unmap_page(struct device *dev, dma_addr_t dma_address, size_t size,
272 enum dma_data_direction direction)
273
274API for mapping and unmapping for pages. All the notes and warnings
275for the other mapping APIs apply here. Also, although the <offset>
276and <size> parameters are provided to do partial page mapping, it is
277recommended that you never use these unless you really know what the
278cache width is.
279
280dma_addr_t
281dma_map_resource(struct device *dev, phys_addr_t phys_addr, size_t size,
282 enum dma_data_direction dir, unsigned long attrs)
283
284void
285dma_unmap_resource(struct device *dev, dma_addr_t addr, size_t size,
286 enum dma_data_direction dir, unsigned long attrs)
287
288API for mapping and unmapping for MMIO resources. All the notes and
289warnings for the other mapping APIs apply here. The API should only be
290used to map device MMIO resources, mapping of RAM is not permitted.
291
292int
293dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
294
295In some circumstances dma_map_single(), dma_map_page() and dma_map_resource()
296will fail to create a mapping. A driver can check for these errors by testing
297the returned DMA address with dma_mapping_error(). A non-zero return value
298means the mapping could not be created and the driver should take appropriate
299action (e.g. reduce current DMA mapping usage or delay and try again later).
300
301 int
302 dma_map_sg(struct device *dev, struct scatterlist *sg,
303 int nents, enum dma_data_direction direction)
304
305Returns: the number of DMA address segments mapped (this may be shorter
306than <nents> passed in if some elements of the scatter/gather list are
307physically or virtually adjacent and an IOMMU maps them with a single
308entry).
309
310Please note that the sg cannot be mapped again if it has been mapped once.
311The mapping process is allowed to destroy information in the sg.
312
313As with the other mapping interfaces, dma_map_sg() can fail. When it
314does, 0 is returned and a driver must take appropriate action. It is
315critical that the driver do something, in the case of a block driver
316aborting the request or even oopsing is better than doing nothing and
317corrupting the filesystem.
318
319With scatterlists, you use the resulting mapping like this:
320
321 int i, count = dma_map_sg(dev, sglist, nents, direction);
322 struct scatterlist *sg;
323
324 for_each_sg(sglist, sg, count, i) {
325 hw_address[i] = sg_dma_address(sg);
326 hw_len[i] = sg_dma_len(sg);
327 }
328
329where nents is the number of entries in the sglist.
330
331The implementation is free to merge several consecutive sglist entries
332into one (e.g. with an IOMMU, or if several pages just happen to be
333physically contiguous) and returns the actual number of sg entries it
334mapped them to. On failure 0, is returned.
335
336Then you should loop count times (note: this can be less than nents times)
337and use sg_dma_address() and sg_dma_len() macros where you previously
338accessed sg->address and sg->length as shown above.
339
340 void
341 dma_unmap_sg(struct device *dev, struct scatterlist *sg,
342 int nents, enum dma_data_direction direction)
343
344Unmap the previously mapped scatter/gather list. All the parameters
345must be the same as those and passed in to the scatter/gather mapping
346API.
347
348Note: <nents> must be the number you passed in, *not* the number of
349DMA address entries returned.
350
351void
352dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle, size_t size,
353 enum dma_data_direction direction)
354void
355dma_sync_single_for_device(struct device *dev, dma_addr_t dma_handle, size_t size,
356 enum dma_data_direction direction)
357void
358dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg, int nents,
359 enum dma_data_direction direction)
360void
361dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg, int nents,
362 enum dma_data_direction direction)
363
364Synchronise a single contiguous or scatter/gather mapping for the CPU
365and device. With the sync_sg API, all the parameters must be the same
366as those passed into the single mapping API. With the sync_single API,
367you can use dma_handle and size parameters that aren't identical to
368those passed into the single mapping API to do a partial sync.
369
370Notes: You must do this:
371
372- Before reading values that have been written by DMA from the device
373 (use the DMA_FROM_DEVICE direction)
374- After writing values that will be written to the device using DMA
375 (use the DMA_TO_DEVICE) direction
376- before *and* after handing memory to the device if the memory is
377 DMA_BIDIRECTIONAL
378
379See also dma_map_single().
380
381dma_addr_t
382dma_map_single_attrs(struct device *dev, void *cpu_addr, size_t size,
383 enum dma_data_direction dir,
384 unsigned long attrs)
385
386void
387dma_unmap_single_attrs(struct device *dev, dma_addr_t dma_addr,
388 size_t size, enum dma_data_direction dir,
389 unsigned long attrs)
390
391int
392dma_map_sg_attrs(struct device *dev, struct scatterlist *sgl,
393 int nents, enum dma_data_direction dir,
394 unsigned long attrs)
395
396void
397dma_unmap_sg_attrs(struct device *dev, struct scatterlist *sgl,
398 int nents, enum dma_data_direction dir,
399 unsigned long attrs)
400
401The four functions above are just like the counterpart functions
402without the _attrs suffixes, except that they pass an optional
403dma_attrs.
404
405The interpretation of DMA attributes is architecture-specific, and
406each attribute should be documented in Documentation/DMA-attributes.txt.
407
408If dma_attrs are 0, the semantics of each of these functions
409is identical to those of the corresponding function
410without the _attrs suffix. As a result dma_map_single_attrs()
411can generally replace dma_map_single(), etc.
412
413As an example of the use of the *_attrs functions, here's how
414you could pass an attribute DMA_ATTR_FOO when mapping memory
415for DMA:
416
417#include <linux/dma-mapping.h>
418/* DMA_ATTR_FOO should be defined in linux/dma-mapping.h and
419 * documented in Documentation/DMA-attributes.txt */
420...
421
422 unsigned long attr;
423 attr |= DMA_ATTR_FOO;
424 ....
425 n = dma_map_sg_attrs(dev, sg, nents, DMA_TO_DEVICE, attr);
426 ....
427
428Architectures that care about DMA_ATTR_FOO would check for its
429presence in their implementations of the mapping and unmapping
430routines, e.g.:
431
432void whizco_dma_map_sg_attrs(struct device *dev, dma_addr_t dma_addr,
433 size_t size, enum dma_data_direction dir,
434 unsigned long attrs)
435{
436 ....
437 if (attrs & DMA_ATTR_FOO)
438 /* twizzle the frobnozzle */
439 ....
440
441
442Part II - Advanced dma_ usage
443-----------------------------
444
445Warning: These pieces of the DMA API should not be used in the
446majority of cases, since they cater for unlikely corner cases that
447don't belong in usual drivers.
448
449If you don't understand how cache line coherency works between a
450processor and an I/O device, you should not be using this part of the
451API at all.
452
453void *
454dma_alloc_noncoherent(struct device *dev, size_t size,
455 dma_addr_t *dma_handle, gfp_t flag)
456
457Identical to dma_alloc_coherent() except that the platform will
458choose to return either consistent or non-consistent memory as it sees
459fit. By using this API, you are guaranteeing to the platform that you
460have all the correct and necessary sync points for this memory in the
461driver should it choose to return non-consistent memory.
462
463Note: where the platform can return consistent memory, it will
464guarantee that the sync points become nops.
465
466Warning: Handling non-consistent memory is a real pain. You should
467only use this API if you positively know your driver will be
468required to work on one of the rare (usually non-PCI) architectures
469that simply cannot make consistent memory.
470
471void
472dma_free_noncoherent(struct device *dev, size_t size, void *cpu_addr,
473 dma_addr_t dma_handle)
474
475Free memory allocated by the nonconsistent API. All parameters must
476be identical to those passed in (and returned by
477dma_alloc_noncoherent()).
478
479int
480dma_get_cache_alignment(void)
481
482Returns the processor cache alignment. This is the absolute minimum
483alignment *and* width that you must observe when either mapping
484memory or doing partial flushes.
485
486Notes: This API may return a number *larger* than the actual cache
487line, but it will guarantee that one or more cache lines fit exactly
488into the width returned by this call. It will also always be a power
489of two for easy alignment.
490
491void
492dma_cache_sync(struct device *dev, void *vaddr, size_t size,
493 enum dma_data_direction direction)
494
495Do a partial sync of memory that was allocated by
496dma_alloc_noncoherent(), starting at virtual address vaddr and
497continuing on for size. Again, you *must* observe the cache line
498boundaries when doing this.
499
500int
501dma_declare_coherent_memory(struct device *dev, phys_addr_t phys_addr,
502 dma_addr_t device_addr, size_t size, int
503 flags)
504
505Declare region of memory to be handed out by dma_alloc_coherent() when
506it's asked for coherent memory for this device.
507
508phys_addr is the CPU physical address to which the memory is currently
509assigned (this will be ioremapped so the CPU can access the region).
510
511device_addr is the DMA address the device needs to be programmed
512with to actually address this memory (this will be handed out as the
513dma_addr_t in dma_alloc_coherent()).
514
515size is the size of the area (must be multiples of PAGE_SIZE).
516
517flags can be ORed together and are:
518
519DMA_MEMORY_MAP - request that the memory returned from
520dma_alloc_coherent() be directly writable.
521
522DMA_MEMORY_IO - request that the memory returned from
523dma_alloc_coherent() be addressable using read()/write()/memcpy_toio() etc.
524
525One or both of these flags must be present.
526
527DMA_MEMORY_INCLUDES_CHILDREN - make the declared memory be allocated by
528dma_alloc_coherent of any child devices of this one (for memory residing
529on a bridge).
530
531DMA_MEMORY_EXCLUSIVE - only allocate memory from the declared regions.
532Do not allow dma_alloc_coherent() to fall back to system memory when
533it's out of memory in the declared region.
534
535The return value will be either DMA_MEMORY_MAP or DMA_MEMORY_IO and
536must correspond to a passed in flag (i.e. no returning DMA_MEMORY_IO
537if only DMA_MEMORY_MAP were passed in) for success or zero for
538failure.
539
540Note, for DMA_MEMORY_IO returns, all subsequent memory returned by
541dma_alloc_coherent() may no longer be accessed directly, but instead
542must be accessed using the correct bus functions. If your driver
543isn't prepared to handle this contingency, it should not specify
544DMA_MEMORY_IO in the input flags.
545
546As a simplification for the platforms, only *one* such region of
547memory may be declared per device.
548
549For reasons of efficiency, most platforms choose to track the declared
550region only at the granularity of a page. For smaller allocations,
551you should use the dma_pool() API.
552
553void
554dma_release_declared_memory(struct device *dev)
555
556Remove the memory region previously declared from the system. This
557API performs *no* in-use checking for this region and will return
558unconditionally having removed all the required structures. It is the
559driver's job to ensure that no parts of this memory region are
560currently in use.
561
562void *
563dma_mark_declared_memory_occupied(struct device *dev,
564 dma_addr_t device_addr, size_t size)
565
566This is used to occupy specific regions of the declared space
567(dma_alloc_coherent() will hand out the first free region it finds).
568
569device_addr is the *device* address of the region requested.
570
571size is the size (and should be a page-sized multiple).
572
573The return value will be either a pointer to the processor virtual
574address of the memory, or an error (via PTR_ERR()) if any part of the
575region is occupied.
576
577Part III - Debug drivers use of the DMA-API
578-------------------------------------------
579
580The DMA-API as described above has some constraints. DMA addresses must be
581released with the corresponding function with the same size for example. With
582the advent of hardware IOMMUs it becomes more and more important that drivers
583do not violate those constraints. In the worst case such a violation can
584result in data corruption up to destroyed filesystems.
585
586To debug drivers and find bugs in the usage of the DMA-API checking code can
587be compiled into the kernel which will tell the developer about those
588violations. If your architecture supports it you can select the "Enable
589debugging of DMA-API usage" option in your kernel configuration. Enabling this
590option has a performance impact. Do not enable it in production kernels.
591
592If you boot the resulting kernel will contain code which does some bookkeeping
593about what DMA memory was allocated for which device. If this code detects an
594error it prints a warning message with some details into your kernel log. An
595example warning message may look like this:
596
597------------[ cut here ]------------
598WARNING: at /data2/repos/linux-2.6-iommu/lib/dma-debug.c:448
599 check_unmap+0x203/0x490()
600Hardware name:
601forcedeth 0000:00:08.0: DMA-API: device driver frees DMA memory with wrong
602 function [device address=0x00000000640444be] [size=66 bytes] [mapped as
603single] [unmapped as page]
604Modules linked in: nfsd exportfs bridge stp llc r8169
605Pid: 0, comm: swapper Tainted: G W 2.6.28-dmatest-09289-g8bb99c0 #1
606Call Trace:
607 <IRQ> [<ffffffff80240b22>] warn_slowpath+0xf2/0x130
608 [<ffffffff80647b70>] _spin_unlock+0x10/0x30
609 [<ffffffff80537e75>] usb_hcd_link_urb_to_ep+0x75/0xc0
610 [<ffffffff80647c22>] _spin_unlock_irqrestore+0x12/0x40
611 [<ffffffff8055347f>] ohci_urb_enqueue+0x19f/0x7c0
612 [<ffffffff80252f96>] queue_work+0x56/0x60
613 [<ffffffff80237e10>] enqueue_task_fair+0x20/0x50
614 [<ffffffff80539279>] usb_hcd_submit_urb+0x379/0xbc0
615 [<ffffffff803b78c3>] cpumask_next_and+0x23/0x40
616 [<ffffffff80235177>] find_busiest_group+0x207/0x8a0
617 [<ffffffff8064784f>] _spin_lock_irqsave+0x1f/0x50
618 [<ffffffff803c7ea3>] check_unmap+0x203/0x490
619 [<ffffffff803c8259>] debug_dma_unmap_page+0x49/0x50
620 [<ffffffff80485f26>] nv_tx_done_optimized+0xc6/0x2c0
621 [<ffffffff80486c13>] nv_nic_irq_optimized+0x73/0x2b0
622 [<ffffffff8026df84>] handle_IRQ_event+0x34/0x70
623 [<ffffffff8026ffe9>] handle_edge_irq+0xc9/0x150
624 [<ffffffff8020e3ab>] do_IRQ+0xcb/0x1c0
625 [<ffffffff8020c093>] ret_from_intr+0x0/0xa
626 <EOI> <4>---[ end trace f6435a98e2a38c0e ]---
627
628The driver developer can find the driver and the device including a stacktrace
629of the DMA-API call which caused this warning.
630
631Per default only the first error will result in a warning message. All other
632errors will only silently counted. This limitation exist to prevent the code
633from flooding your kernel log. To support debugging a device driver this can
634be disabled via debugfs. See the debugfs interface documentation below for
635details.
636
637The debugfs directory for the DMA-API debugging code is called dma-api/. In
638this directory the following files can currently be found:
639
640 dma-api/all_errors This file contains a numeric value. If this
641 value is not equal to zero the debugging code
642 will print a warning for every error it finds
643 into the kernel log. Be careful with this
644 option, as it can easily flood your logs.
645
646 dma-api/disabled This read-only file contains the character 'Y'
647 if the debugging code is disabled. This can
648 happen when it runs out of memory or if it was
649 disabled at boot time
650
651 dma-api/error_count This file is read-only and shows the total
652 numbers of errors found.
653
654 dma-api/num_errors The number in this file shows how many
655 warnings will be printed to the kernel log
656 before it stops. This number is initialized to
657 one at system boot and be set by writing into
658 this file
659
660 dma-api/min_free_entries
661 This read-only file can be read to get the
662 minimum number of free dma_debug_entries the
663 allocator has ever seen. If this value goes
664 down to zero the code will disable itself
665 because it is not longer reliable.
666
667 dma-api/num_free_entries
668 The current number of free dma_debug_entries
669 in the allocator.
670
671 dma-api/driver-filter
672 You can write a name of a driver into this file
673 to limit the debug output to requests from that
674 particular driver. Write an empty string to
675 that file to disable the filter and see
676 all errors again.
677
678If you have this code compiled into your kernel it will be enabled by default.
679If you want to boot without the bookkeeping anyway you can provide
680'dma_debug=off' as a boot parameter. This will disable DMA-API debugging.
681Notice that you can not enable it again at runtime. You have to reboot to do
682so.
683
684If you want to see debug messages only for a special device driver you can
685specify the dma_debug_driver=<drivername> parameter. This will enable the
686driver filter at boot time. The debug code will only print errors for that
687driver afterwards. This filter can be disabled or changed later using debugfs.
688
689When the code disables itself at runtime this is most likely because it ran
690out of dma_debug_entries. These entries are preallocated at boot. The number
691of preallocated entries is defined per architecture. If it is too low for you
692boot with 'dma_debug_entries=<your_desired_number>' to overwrite the
693architectural default.
694
695void debug_dmap_mapping_error(struct device *dev, dma_addr_t dma_addr);
696
697dma-debug interface debug_dma_mapping_error() to debug drivers that fail
698to check DMA mapping errors on addresses returned by dma_map_single() and
699dma_map_page() interfaces. This interface clears a flag set by
700debug_dma_map_page() to indicate that dma_mapping_error() has been called by
701the driver. When driver does unmap, debug_dma_unmap() checks the flag and if
702this flag is still set, prints warning message that includes call trace that
703leads up to the unmap. This interface can be called from dma_mapping_error()
704routines to enable DMA mapping error check debugging.
705