xref: /freebsd/share/man/man9/bus_dma.9 (revision 6966ac055c3b7a39266fb982493330df7a097997)
1.\" Copyright (c) 2002, 2003 Hiten M. Pandya.
2.\" All rights reserved.
3.\"
4.\" Redistribution and use in source and binary forms, with or without
5.\" modification, are permitted provided that the following conditions
6.\" are met:
7.\" 1. Redistributions of source code must retain the above copyright
8.\"    notice, this list of conditions, and the following disclaimer,
9.\"    without modification, immediately at the beginning of the file.
10.\" 2. The name of the author may not be used to endorse or promote products
11.\"    derived from this software without specific prior written permission.
12.\"
13.\" THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14.\" ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15.\" IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16.\" ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR, CONTRIBUTORS OR THE
17.\" VOICES IN HITEN PANDYA'S HEAD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
18.\" SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
19.\" TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
20.\" PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
21.\" LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
22.\" NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
23.\" SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
24.\"
25.\" Copyright (c) 1996, 1997, 1998, 2001 The NetBSD Foundation, Inc.
26.\" All rights reserved.
27.\"
28.\" This code is derived from software contributed to The NetBSD Foundation
29.\" by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
30.\" NASA Ames Research Center.
31.\"
32.\" Redistribution and use in source and binary forms, with or without
33.\" modification, are permitted provided that the following conditions
34.\" are met:
35.\" 1. Redistributions of source code must retain the above copyright
36.\"    notice, this list of conditions and the following disclaimer.
37.\" 2. Redistributions in binary form must reproduce the above copyright
38.\"    notice, this list of conditions and the following disclaimer in the
39.\"    documentation and/or other materials provided with the distribution.
40.\"
41.\" THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
42.\" ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
43.\" TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
44.\" PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
45.\" BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
46.\" CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
47.\" SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
48.\" INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
49.\" CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
50.\" ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
51.\" POSSIBILITY OF SUCH DAMAGE.
52.\"
53.\" $FreeBSD$
54.\" $NetBSD: bus_dma.9,v 1.25 2002/10/14 13:43:16 wiz Exp $
55.\"
56.Dd August 11, 2018
57.Dt BUS_DMA 9
58.Os
59.Sh NAME
60.Nm bus_dma ,
61.Nm bus_dma_tag_create ,
62.Nm bus_dma_tag_destroy ,
63.Nm bus_dma_template_init ,
64.Nm bus_dma_template_tag ,
65.Nm bus_dma_template_clone ,
66.Nm bus_dmamap_create ,
67.Nm bus_dmamap_destroy ,
68.Nm bus_dmamap_load ,
69.Nm bus_dmamap_load_bio ,
70.Nm bus_dmamap_load_ccb ,
71.Nm bus_dmamap_load_mbuf ,
72.Nm bus_dmamap_load_mbuf_sg ,
73.Nm bus_dmamap_load_uio ,
74.Nm bus_dmamap_unload ,
75.Nm bus_dmamap_sync ,
76.Nm bus_dmamem_alloc ,
77.Nm bus_dmamem_free
78.Nd Bus and Machine Independent DMA Mapping Interface
79.Sh SYNOPSIS
80.In machine/bus.h
81.Ft int
82.Fn bus_dma_tag_create "bus_dma_tag_t parent" "bus_size_t alignment" \
83"bus_addr_t boundary" "bus_addr_t lowaddr" "bus_addr_t highaddr" \
84"bus_dma_filter_t *filtfunc" "void *filtfuncarg" "bus_size_t maxsize" \
85"int nsegments" "bus_size_t maxsegsz" "int flags" "bus_dma_lock_t *lockfunc" \
86"void *lockfuncarg" "bus_dma_tag_t *dmat"
87.Ft int
88.Fn bus_dma_tag_destroy "bus_dma_tag_t dmat"
89.Ft void
90.Fo bus_dma_template_init
91.Fa "bus_dma_template_t template"
92.Fa "bus_dma_tag_t parent"
93.Fc
94.Ft int
95.Fo bus_dma_template_tag
96.Fa "bus_dma_template_t template"
97.Fa "bus_dma_tag_t *dmat"
98.Fc
99.Ft void
100.Fo bus_dma_template_clone
101.Fa "bus_dma_template_t template"
102.Fa "bus_dma_tag_t dmat"
103.Fc
104.Ft int
105.Fn bus_dmamap_create "bus_dma_tag_t dmat" "int flags" "bus_dmamap_t *mapp"
106.Ft int
107.Fn bus_dmamap_destroy "bus_dma_tag_t dmat" "bus_dmamap_t map"
108.Ft int
109.Fn bus_dmamap_load "bus_dma_tag_t dmat" "bus_dmamap_t map" "void *buf" \
110"bus_size_t buflen" "bus_dmamap_callback_t *callback" "void *callback_arg" \
111"int flags"
112.Ft int
113.Fn bus_dmamap_load_bio "bus_dma_tag_t dmat" "bus_dmamap_t map" \
114"struct bio *bio" "bus_dmamap_callback_t *callback" "void *callback_arg" \
115"int flags"
116.Ft int
117.Fn bus_dmamap_load_ccb "bus_dma_tag_t dmat" "bus_dmamap_t map" \
118"union ccb *ccb" "bus_dmamap_callback_t *callback" "void *callback_arg" \
119"int flags"
120.Ft int
121.Fn bus_dmamap_load_mbuf "bus_dma_tag_t dmat" "bus_dmamap_t map" \
122"struct mbuf *mbuf" "bus_dmamap_callback2_t *callback" "void *callback_arg" \
123"int flags"
124.Ft int
125.Fn bus_dmamap_load_mbuf_sg "bus_dma_tag_t dmat" "bus_dmamap_t map" \
126"struct mbuf *mbuf" "bus_dma_segment_t *segs" "int *nsegs" "int flags"
127.Ft int
128.Fn bus_dmamap_load_uio "bus_dma_tag_t dmat" "bus_dmamap_t map" \
129"struct uio *uio" "bus_dmamap_callback2_t *callback" "void *callback_arg" \
130"int flags"
131.Ft void
132.Fn bus_dmamap_unload "bus_dma_tag_t dmat" "bus_dmamap_t map"
133.Ft void
134.Fn bus_dmamap_sync "bus_dma_tag_t dmat" "bus_dmamap_t map" \
135"op"
136.Ft int
137.Fn bus_dmamem_alloc "bus_dma_tag_t dmat" "void **vaddr" \
138"int flags" "bus_dmamap_t *mapp"
139.Ft void
140.Fn bus_dmamem_free "bus_dma_tag_t dmat" "void *vaddr" \
141"bus_dmamap_t map"
142.Sh DESCRIPTION
143Direct Memory Access (DMA) is a method of transferring data
144without involving the CPU, thus providing higher performance.
145A DMA transaction can be achieved between device to memory,
146device to device, or memory to memory.
147.Pp
148The
149.Nm
150API is a bus, device, and machine-independent (MI) interface to
151DMA mechanisms.
152It provides the client with flexibility and simplicity by
153abstracting machine dependent issues like setting up
154DMA mappings, handling cache issues, bus specific features
155and limitations.
156.Sh OVERVIEW
157A tag structure
158.Vt ( bus_dma_tag_t )
159is used to describe the properties of a group of related DMA
160transactions.
161One way to view this is that a tag describes the limitations of a DMA engine.
162For example, if a DMA engine in a device is limited to 32-bit addresses,
163that limitation is specified by a parameter when creating the tag
164for that device.
165Similarly, a tag can be marked as requiring buffers whose addresses are
166aligned to a specific boundary.
167.Pp
168Some devices may require multiple tags to describe DMA
169transactions with differing properties.
170For example, a device might require 16-byte alignment of its descriptor ring
171while permitting arbitrary alignment of I/O buffers.
172In this case,
173the driver must create one tag for the descriptor ring and a separate tag for
174I/O buffers.
175If a device has restrictions that are common to all DMA transactions
176in addition to restrictions that differ between unrelated groups of
177transactions,
178the driver can first create a
179.Dq parent
180tag that decribes the common restrictions.
181The per-group tags can then inherit these restrictions from this
182.Dq parent
183tag rather than having to list them explicitly when creating the per-group tags.
184.Pp
185A mapping structure
186.Vt ( bus_dmamap_t )
187represents a mapping of a memory region for DMA.
188On systems with I/O MMUs,
189the mapping structure tracks any I/O MMU entries used by a request.
190For DMA requests that require bounce pages,
191the mapping tracks the bounce pages used.
192.Pp
193To prepare for one or more DMA transactions,
194a mapping must be bound to a memory region by calling one of the
195.Fn bus_dmamap_load
196functions.
197These functions configure the mapping which can include programming entries
198in an I/O MMU and/or allocating bounce pages.
199An output of these functions
200(either directly or indirectly by invoking a callback routine)
201is the list of scatter/gather address ranges a consumer can pass to a DMA
202engine to access the memory region.
203When a mapping is no longer needed,
204the mapping must be unloaded via
205.Fn bus_dmamap_unload .
206.Pp
207Before and after each DMA transaction,
208.Fn bus_dmamap_sync
209must be used to ensure that the correct data is used by the DMA engine and
210the CPU.
211If a mapping uses bounce pages,
212the sync operations copy data between the bounce pages and the memory region
213bound to the mapping.
214Sync operations also handle architecture-specific details such as CPU cache
215flushing and CPU memory operation ordering.
216.Sh STATIC VS DYNAMIC
217.Nm
218handles two types of DMA transactions: static and dynamic.
219Static transactions are used with a long-lived memory region that is reused
220for many transactions such as a descriptor ring.
221Dynamic transactions are used for transfers to or from transient buffers
222such as I/O buffers holding a network packet or disk block.
223Each transaction type uses a different subset of the
224.Nm
225API.
226.Ss Static Transactions
227Static transactions use memory regions allocated by
228.Nm .
229Each static memory region is allocated by calling
230.Fn bus_dmamem_alloc .
231This function requires a valid tag describing the properties of the
232DMA transactions to this region such as alignment or address restrictions.
233Multiple regions can share a single tag if they share the same restrictions.
234.Pp
235.Fn bus_dmamem_alloc
236allocates a memory region along with a mapping object.
237The associated tag, memory region, and mapping object must then be passed to
238.Fn bus_dmamap_load
239to bind the mapping to the allocated region and obtain the
240scatter/gather list.
241.Pp
242It is expected that
243.Fn bus_dmamem_alloc
244will attempt to allocate memory requiring less expensive sync operations
245(for example, implementations should not allocate regions requiring bounce
246pages),
247but sync operations should still be used.
248For example, a driver should use
249.Fn bus_dmamap_sync
250in an interrupt handler before reading descriptor ring entries written by the
251device prior to the interrupt.
252.Pp
253When a consumer is finished with a memory region,
254it should unload the mapping via
255.Fn bus_dmamap_unload
256and then release the memory region and mapping object via
257.Fn bus_dmamem_free .
258.Ss Dynamic Transactions
259Dynamic transactions map memory regions provided by other parts of the system.
260A tag must be created via
261.Fn bus_dma_tag_create
262to describe the DMA transactions to and from these memory regions,
263and a pool of mapping objects must be allocated via
264.Fn bus_dmamap_create
265to track the mappings of any in-flight transactions.
266.Pp
267When a consumer wishes to schedule a transaction for a memory region,
268the consumer must first obtain an unused mapping object from its pool
269of mapping objects.
270The memory region must be bound to the mapping object via one of the
271.Fn bus_dmamap_load
272functions.
273Before scheduling the transaction,
274the consumer should sync the memory region via
275.Fn bus_dmamap_sync
276with one or more of the
277.Dq PRE
278flags.
279After the transaction has completed,
280the consumer should sync the memory region via
281.Fn bus_dmamap_sync
282with one or more of the
283.Dq POST
284flags.
285The mapping can then be unloaded via
286.Fn bus_dmamap_unload ,
287and the mapping object can be returned to the pool of unused mapping objects.
288.Pp
289When a consumer is no longer scheduling DMA transactions,
290the mapping objects should be freed via
291.Fn bus_dmamap_destroy ,
292and the tag should be freed via
293.Fn bus_dma_tag_destroy .
294.Sh STRUCTURES AND TYPES
295.Bl -tag -width indent
296.It Vt bus_dma_tag_t
297A machine-dependent (MD) opaque type that describes the
298characteristics of a group of DMA transactions.
299DMA tags are organized into a hierarchy, with each child
300tag inheriting the restrictions of its parent.
301This allows all devices along the path of DMA transactions
302to contribute to the constraints of those transactions.
303.It Vt bus_dma_template_t
304A template structure for creating a
305.Fa bus_dma_tag_t
306from a set of defaults.
307Once initialized with
308.Fn bus_dma_template_init ,
309a driver can over-ride individual fields to suit its needs.
310The following fields have the indicated values:
311.Bd -literal
312	alignment	1
313	boundary	0
314	lowaddr		BUS_SPACE_MAXADDR
315	highaddr	BUS_SPACE_MAXADDR
316	maxsize		BUS_SPACE_MAXSIZE
317	nsegments	BUS_SPACE_UNRESTRICTED
318	maxsegsize	BUS_SPACE_MAXSIZE
319	flags		0
320	lockfunc	NULL
321	lockfuncarg	NULL
322.Ed
323.Pp
324Descriptions of each field are documented with
325.Fn bus_dma_tag_create .
326Note that the
327.Fa filtfunc
328and
329.Fa filtfuncarg
330attributes of the DMA tag are not supported with templates.
331.It Vt bus_dma_filter_t
332Client specified address filter having the format:
333.Bl -tag -width indent
334.It Ft int
335.Fn "client_filter" "void *filtarg" "bus_addr_t testaddr"
336.El
337.Pp
338Address filters can be specified during tag creation to allow
339for devices whose DMA address restrictions cannot be specified
340by a single window.
341The
342.Fa filtarg
343argument is specified by the client during tag creation to be passed to all
344invocations of the callback.
345The
346.Fa testaddr
347argument contains a potential starting address of a DMA mapping.
348The filter function operates on the set of addresses from
349.Fa testaddr
350to
351.Ql trunc_page(testaddr) + PAGE_SIZE - 1 ,
352inclusive.
353The filter function should return zero if any mapping in this range
354can be accommodated by the device and non-zero otherwise.
355.It Vt bus_dma_segment_t
356A machine-dependent type that describes individual
357DMA segments.
358It contains the following fields:
359.Bd -literal
360	bus_addr_t	ds_addr;
361	bus_size_t	ds_len;
362.Ed
363.Pp
364The
365.Fa ds_addr
366field contains the device visible address of the DMA segment, and
367.Fa ds_len
368contains the length of the DMA segment.
369Although the DMA segments returned by a mapping call will adhere to
370all restrictions necessary for a successful DMA operation, some conversion
371(e.g.\& a conversion from host byte order to the device's byte order) is
372almost always required when presenting segment information to the device.
373.It Vt bus_dmamap_t
374A machine-dependent opaque type describing an individual mapping.
375One map is used for each memory allocation that will be loaded.
376Maps can be reused once they have been unloaded.
377Multiple maps can be associated with one DMA tag.
378While the value of the map may evaluate to
379.Dv NULL
380on some platforms under certain conditions,
381it should never be assumed that it will be
382.Dv NULL
383in all cases.
384.It Vt bus_dmamap_callback_t
385Client specified callback for receiving mapping information resulting from
386the load of a
387.Vt bus_dmamap_t
388via
389.Fn bus_dmamap_load ,
390.Fn bus_dmamap_load_bio
391or
392.Fn bus_dmamap_load_ccb .
393Callbacks are of the format:
394.Bl -tag -width indent
395.It Ft void
396.Fn "client_callback" "void *callback_arg" "bus_dma_segment_t *segs" \
397"int nseg" "int error"
398.El
399.Pp
400The
401.Fa callback_arg
402is the callback argument passed to dmamap load functions.
403The
404.Fa segs
405and
406.Fa nseg
407arguments describe an array of
408.Vt bus_dma_segment_t
409structures that represent the mapping.
410This array is only valid within the scope of the callback function.
411The success or failure of the mapping is indicated by the
412.Fa error
413argument.
414More information on the use of callbacks can be found in the
415description of the individual dmamap load functions.
416.It Vt bus_dmamap_callback2_t
417Client specified callback for receiving mapping information resulting from
418the load of a
419.Vt bus_dmamap_t
420via
421.Fn bus_dmamap_load_uio
422or
423.Fn bus_dmamap_load_mbuf .
424.Pp
425Callback2s are of the format:
426.Bl -tag -width indent
427.It Ft void
428.Fn "client_callback2" "void *callback_arg" "bus_dma_segment_t *segs" \
429"int nseg" "bus_size_t mapsize" "int error"
430.El
431.Pp
432Callback2's behavior is the same as
433.Vt bus_dmamap_callback_t
434with the addition that the length of the data mapped is provided via
435.Fa mapsize .
436.It Vt bus_dmasync_op_t
437Memory synchronization operation specifier.
438Bus DMA requires explicit synchronization of memory with its device
439visible mapping in order to guarantee memory coherency.
440The
441.Vt bus_dmasync_op_t
442allows the type of DMA operation that will be or has been performed
443to be communicated to the system so that the correct coherency measures
444are taken.
445The operations are represented as bitfield flags that can be combined together,
446though it only makes sense to combine PRE flags or POST flags, not both.
447See the
448.Fn bus_dmamap_sync
449description below for more details on how to use these operations.
450.Pp
451All operations specified below are performed from the host memory point of view,
452where a read implies data coming from the device to the host memory, and a write
453implies data going from the host memory to the device.
454Alternatively, the operations can be thought of in terms of driver operations,
455where reading a network packet or storage sector corresponds to a read operation
456in
457.Nm .
458.Bl -tag -width ".Dv BUS_DMASYNC_POSTWRITE"
459.It Dv BUS_DMASYNC_PREREAD
460Perform any synchronization required prior to an update of host memory by the
461device.
462.It Dv BUS_DMASYNC_PREWRITE
463Perform any synchronization required after an update of host memory by the CPU
464and prior to device access to host memory.
465.It Dv BUS_DMASYNC_POSTREAD
466Perform any synchronization required after an update of host memory by the
467device and prior to CPU access to host memory.
468.It Dv BUS_DMASYNC_POSTWRITE
469Perform any synchronization required after device access to host memory.
470.El
471.It Vt bus_dma_lock_t
472Client specified lock/mutex manipulation method.
473This will be called from
474within busdma whenever a client lock needs to be manipulated.
475In its current form, the function will be called immediately before
476the callback for a DMA load operation that has been deferred with
477.Dv BUS_DMA_LOCK
478and immediately after with
479.Dv BUS_DMA_UNLOCK .
480If the load operation does not need to be deferred, then it
481will not be called since the function loading the map should
482be holding the appropriate locks.
483This method is of the format:
484.Bl -tag -width indent
485.It Ft void
486.Fn "lockfunc" "void *lockfunc_arg" "bus_dma_lock_op_t op"
487.El
488.Pp
489The
490.Fa lockfuncarg
491argument is specified by the client during tag creation to be passed to all
492invocations of the callback.
493The
494.Fa op
495argument specifies the lock operation to perform.
496.Pp
497Two
498.Vt lockfunc
499implementations are provided for convenience.
500.Fn busdma_lock_mutex
501performs standard mutex operations on the sleep mutex provided via
502.Fa lockfuncarg .
503.Fn dflt_lock
504will generate a system panic if it is called.
505It is substituted into the tag when
506.Fa lockfunc
507is passed as
508.Dv NULL
509to
510.Fn bus_dma_tag_create
511and is useful for tags that should not be used with deferred load operations.
512.It Vt bus_dma_lock_op_t
513Operations to be performed by the client-specified
514.Fn lockfunc .
515.Bl -tag -width ".Dv BUS_DMA_UNLOCK"
516.It Dv BUS_DMA_LOCK
517Acquires and/or locks the client locking primitive.
518.It Dv BUS_DMA_UNLOCK
519Releases and/or unlocks the client locking primitive.
520.El
521.El
522.Sh FUNCTIONS
523.Bl -tag -width indent
524.It Fn bus_dma_tag_create "parent" "alignment" "boundary" "lowaddr" \
525"highaddr" "*filtfunc" "*filtfuncarg" "maxsize" "nsegments" "maxsegsz" \
526"flags" "lockfunc" "lockfuncarg" "*dmat"
527Allocates a DMA tag, and initializes it according to
528the arguments provided:
529.Bl -tag -width ".Fa filtfuncarg"
530.It Fa parent
531A parent tag from which to inherit restrictions.
532The restrictions passed in other arguments can only further tighten the
533restrictions inherited from the parent tag.
534.Pp
535All tags created by a device driver must inherit from the tag returned by
536.Fn bus_get_dma_tag
537to honor restrictions between the parent bridge, CPU memory, and the
538device.
539.It Fa alignment
540Alignment constraint, in bytes, of any mappings created using this tag.
541The alignment must be a power of 2.
542Hardware that can DMA starting at any address would specify
543.Em 1
544for byte alignment.
545Hardware requiring DMA transfers to start on a multiple of 4K
546would specify
547.Em 4096 .
548.It Fa boundary
549Boundary constraint, in bytes, of the target DMA memory region.
550The boundary indicates the set of addresses, all multiples of the
551boundary argument, that cannot be crossed by a single
552.Vt bus_dma_segment_t .
553The boundary must be a power of 2 and must be no smaller than the
554maximum segment size.
555.Ql 0
556indicates that there are no boundary restrictions.
557.It Fa lowaddr , highaddr
558Bounds of the window of bus address space that
559.Em cannot
560be directly accessed by the device.
561The window contains all addresses greater than
562.Fa lowaddr
563and less than or equal to
564.Fa highaddr .
565For example, a device incapable of DMA above 4GB, would specify a
566.Fa highaddr
567of
568.Dv BUS_SPACE_MAXADDR
569and a
570.Fa lowaddr
571of
572.Dv BUS_SPACE_MAXADDR_32BIT .
573Similarly a device that can only perform DMA to addresses below
57416MB would specify a
575.Fa highaddr
576of
577.Dv BUS_SPACE_MAXADDR
578and a
579.Fa lowaddr
580of
581.Dv BUS_SPACE_MAXADDR_24BIT .
582Some implementations require that some region of device visible
583address space, overlapping available host memory, be outside the
584window.
585This area of
586.Ql safe memory
587is used to bounce requests that would otherwise conflict with
588the exclusion window.
589.It Fa filtfunc
590Optional filter function (may be
591.Dv NULL )
592to be called for any attempt to
593map memory into the window described by
594.Fa lowaddr
595and
596.Fa highaddr .
597A filter function is only required when the single window described
598by
599.Fa lowaddr
600and
601.Fa highaddr
602cannot adequately describe the constraints of the device.
603The filter function will be called for every machine page
604that overlaps the exclusion window.
605.It Fa filtfuncarg
606Argument passed to all calls to the filter function for this tag.
607May be
608.Dv NULL .
609.It Fa maxsize
610Maximum size, in bytes, of the sum of all segment lengths in a given
611DMA mapping associated with this tag.
612.It Fa nsegments
613Number of discontinuities (scatter/gather segments) allowed
614in a DMA mapped region.
615If there is no restriction,
616.Dv BUS_SPACE_UNRESTRICTED
617may be specified.
618.It Fa maxsegsz
619Maximum size, in bytes, of a segment in any DMA mapped region associated
620with
621.Fa dmat .
622.It Fa flags
623Are as follows:
624.Bl -tag -width ".Dv BUS_DMA_ALLOCNOW"
625.It Dv BUS_DMA_ALLOCNOW
626Pre-allocate enough resources to handle at least one map load operation on
627this tag.
628If sufficient resources are not available,
629.Er ENOMEM
630is returned.
631This should not be used for tags that only describe buffers that will be
632allocated with
633.Fn bus_dmamem_alloc .
634Also, due to resource sharing with other tags, this flag does not guarantee
635that resources will be allocated or reserved exclusively for this tag.
636It should be treated only as a minor optimization.
637.It Dv BUS_DMA_COHERENT
638Indicate that the DMA engine and CPU are cache-coherent.
639Cached memory may be used to back allocations created by
640.Fn bus_dmamem_alloc .
641For
642.Fn bus_dma_tag_create ,
643the
644.Dv BUS_DMA_COHERENT
645flag is currently implemented on arm64.
646.El
647.It Fa lockfunc
648Optional lock manipulation function (may be
649.Dv NULL )
650to be called when busdma
651needs to manipulate a lock on behalf of the client.
652If
653.Dv NULL
654is specified,
655.Fn dflt_lock
656is used.
657.It Fa lockfuncarg
658Optional argument to be passed to the function specified by
659.Fa lockfunc .
660.It Fa dmat
661Pointer to a bus_dma_tag_t where the resulting DMA tag will
662be stored.
663.El
664.Pp
665Returns
666.Er ENOMEM
667if sufficient memory is not available for tag creation
668or allocating mapping resources.
669.It Fn bus_dma_tag_destroy "dmat"
670Deallocate the DMA tag
671.Fa dmat
672that was created by
673.Fn bus_dma_tag_create .
674.Pp
675Returns
676.Er EBUSY
677if any DMA maps remain associated with
678.Fa dmat
679or
680.Ql 0
681on success.
682.It Fn bus_dma_template_init "*template" "parent"
683Initializes a
684.Fa bus_dma_template_t
685structure and associates it with an optional
686.Fa parent .
687The
688.Fa parent
689argument may be NULL.
690.It Fn bus_dma_template_tag "*template" "*dmat"
691Unpacks a template into a tag, and returns the tag via the
692.Fa dmat .
693All return values are identical to
694.Fn bus_dma_tag_create .
695.It Fn bus_dma_template_clone "*template" "dmat"
696Clones the fields from a tag to a template.
697This is useful for cloning tags when paired with
698.Fn bus_dma_template_tag .
699A template that is filled in as a clone does not need to be initialized
700first.
701.It Fn bus_dmamap_create "dmat" "flags" "*mapp"
702Allocates and initializes a DMA map.
703Arguments are as follows:
704.Bl -tag -width ".Fa nsegments"
705.It Fa dmat
706DMA tag.
707.It Fa flags
708Are as follows:
709.Bl -tag -width ".Dv BUS_DMA_COHERENT"
710.It Dv BUS_DMA_COHERENT
711Attempt to map the memory loaded with this map such that cache sync
712operations are as cheap as possible.
713This flag is typically set on maps when the memory loaded with these will
714be accessed by both a CPU and a DMA engine, frequently such as control data
715and as opposed to streamable data such as receive and transmit buffers.
716Use of this flag does not remove the requirement of using
717.Fn bus_dmamap_sync ,
718but it may reduce the cost of performing these operations.
719For
720.Fn bus_dmamap_create ,
721the
722.Dv BUS_DMA_COHERENT
723flag is currently implemented on sparc64.
724.El
725.It Fa mapp
726Pointer to a
727.Vt bus_dmamap_t
728where the resulting DMA map will be stored.
729.El
730.Pp
731Returns
732.Er ENOMEM
733if sufficient memory is not available for creating the
734map or allocating mapping resources.
735.It Fn bus_dmamap_destroy "dmat" "map"
736Frees all resources associated with a given DMA map.
737Arguments are as follows:
738.Bl -tag -width ".Fa dmat"
739.It Fa dmat
740DMA tag used to allocate
741.Fa map .
742.It Fa map
743The DMA map to destroy.
744.El
745.Pp
746Returns
747.Er EBUSY
748if a mapping is still active for
749.Fa map .
750.It Fn bus_dmamap_load "dmat" "map" "buf" "buflen" "*callback" \
751"callback_arg" "flags"
752Creates a mapping in device visible address space of
753.Fa buflen
754bytes of
755.Fa buf ,
756associated with the DMA map
757.Fa map .
758This call will always return immediately and will not block for any reason.
759Arguments are as follows:
760.Bl -tag -width ".Fa buflen"
761.It Fa dmat
762DMA tag used to allocate
763.Fa map .
764.It Fa map
765A DMA map without a currently active mapping.
766.It Fa buf
767A kernel virtual address pointer to a contiguous (in KVA) buffer, to be
768mapped into device visible address space.
769.It Fa buflen
770The size of the buffer.
771.It Fa callback Fa callback_arg
772The callback function, and its argument.
773This function is called once sufficient mapping resources are available for
774the DMA operation.
775If resources are temporarily unavailable, this function will be deferred until
776later, but the load operation will still return immediately to the caller.
777Thus, callers should not assume that the callback will be called before the
778load returns, and code should be structured appropriately to handle this.
779See below for specific flags and error codes that control this behavior.
780.It Fa flags
781Are as follows:
782.Bl -tag -width ".Dv BUS_DMA_NOWAIT"
783.It Dv BUS_DMA_NOWAIT
784The load should not be deferred in case of insufficient mapping resources,
785and instead should return immediately with an appropriate error.
786.It Dv BUS_DMA_NOCACHE
787The generated transactions to and from the virtual page are non-cacheable.
788For
789.Fn bus_dmamap_load ,
790the
791.Dv BUS_DMA_NOCACHE
792flag is currently implemented on sparc64.
793.El
794.El
795.Pp
796Return values to the caller are as follows:
797.Bl -tag -width ".Er EINPROGRESS"
798.It 0
799The callback has been called and completed.
800The status of the mapping has been delivered to the callback.
801.It Er EINPROGRESS
802The mapping has been deferred for lack of resources.
803The callback will be called as soon as resources are available.
804Callbacks are serviced in FIFO order.
805.Pp
806Note that subsequent load operations for the same tag that do not require
807extra resources will still succeed.
808This may result in out-of-order processing of requests.
809If the caller requires the order of requests to be preserved,
810then the caller is required to stall subsequent requests until a pending
811request's callback is invoked.
812.It Er ENOMEM
813The load request has failed due to insufficient resources, and the caller
814specifically used the
815.Dv BUS_DMA_NOWAIT
816flag.
817.It Er EINVAL
818The load request was invalid.
819The callback has been called and has been provided the same error.
820This error value may indicate that
821.Fa dmat ,
822.Fa map ,
823.Fa buf ,
824or
825.Fa callback
826were invalid, or
827.Fa buflen
828was larger than the
829.Fa maxsize
830argument used to create the dma tag
831.Fa dmat .
832.El
833.Pp
834When the callback is called, it is presented with an error value
835indicating the disposition of the mapping.
836Error may be one of the following:
837.Bl -tag -width ".Er EINPROGRESS"
838.It 0
839The mapping was successful and the
840.Fa dm_segs
841callback argument contains an array of
842.Vt bus_dma_segment_t
843elements describing the mapping.
844This array is only valid during the scope of the callback function.
845.It Er EFBIG
846A mapping could not be achieved within the segment constraints provided
847in the tag even though the requested allocation size was less than maxsize.
848.El
849.It Fn bus_dmamap_load_bio "dmat" "map" "bio" "callback" "callback_arg" "flags"
850This is a variation of
851.Fn bus_dmamap_load
852which maps buffers pointed to by
853.Fa bio
854for DMA transfers.
855.Fa bio
856may point to either a mapped or unmapped buffer.
857.It Fn bus_dmamap_load_ccb "dmat" "map" "ccb" "callback" "callback_arg" "flags"
858This is a variation of
859.Fn bus_dmamap_load
860which maps data pointed to by
861.Fa ccb
862for DMA transfers.
863The data for
864.Fa ccb
865may be any of the following types:
866.Bl -tag -width ".Er CAM_DATA_SG_PADDR"
867.It CAM_DATA_VADDR
868The data is a single KVA buffer.
869.It CAM_DATA_PADDR
870The data is a single bus address range.
871.It CAM_DATA_SG
872The data is a scatter/gather list of KVA buffers.
873.It CAM_DATA_SG_PADDR
874The data is a scatter/gather list of bus address ranges.
875.It CAM_DATA_BIO
876The data is contained in a
877.Vt struct bio
878attached to the CCB.
879.El
880.Pp
881.Fn bus_dmamap_load_ccb
882supports the following CCB XPT function codes:
883.Pp
884.Bl -item -offset indent -compact
885.It
886XPT_ATA_IO
887.It
888XPT_CONT_TARGET_IO
889.It
890XPT_SCSI_IO
891.El
892.It Fn bus_dmamap_load_mbuf "dmat" "map" "mbuf" "callback2" "callback_arg" \
893"flags"
894This is a variation of
895.Fn bus_dmamap_load
896which maps mbuf chains
897for DMA transfers.
898A
899.Vt bus_size_t
900argument is also passed to the callback routine, which
901contains the mbuf chain's packet header length.
902The
903.Dv BUS_DMA_NOWAIT
904flag is implied, thus no callback deferral will happen.
905.Pp
906Mbuf chains are assumed to be in kernel virtual address space.
907.Pp
908Beside the error values listed for
909.Fn bus_dmamap_load ,
910.Er EINVAL
911will be returned if the size of the mbuf chain exceeds the maximum limit of the
912DMA tag.
913.It Fn bus_dmamap_load_mbuf_sg "dmat" "map" "mbuf" "segs" "nsegs" "flags"
914This is just like
915.Fn bus_dmamap_load_mbuf
916except that it returns immediately without calling a callback function.
917It is provided for efficiency.
918The scatter/gather segment array
919.Va segs
920is provided by the caller and filled in directly by the function.
921The
922.Va nsegs
923argument is returned with the number of segments filled in.
924Returns the same errors as
925.Fn bus_dmamap_load_mbuf .
926.It Fn bus_dmamap_load_uio "dmat" "map" "uio" "callback2" "callback_arg" "flags"
927This is a variation of
928.Fn bus_dmamap_load
929which maps buffers pointed to by
930.Fa uio
931for DMA transfers.
932A
933.Vt bus_size_t
934argument is also passed to the callback routine, which contains the size of
935.Fa uio ,
936i.e.
937.Fa uio->uio_resid .
938The
939.Dv BUS_DMA_NOWAIT
940flag is implied, thus no callback deferral will happen.
941Returns the same errors as
942.Fn bus_dmamap_load .
943.Pp
944If
945.Fa uio->uio_segflg
946is
947.Dv UIO_USERSPACE ,
948then it is assumed that the buffer,
949.Fa uio
950is in
951.Fa "uio->uio_td->td_proc" Ns 's
952address space.
953User space memory must be in-core and wired prior to attempting a map
954load operation.
955Pages may be locked using
956.Xr vslock 9 .
957.It Fn bus_dmamap_unload "dmat" "map"
958Unloads a DMA map.
959Arguments are as follows:
960.Bl -tag -width ".Fa dmam"
961.It Fa dmat
962DMA tag used to allocate
963.Fa map .
964.It Fa map
965The DMA map that is to be unloaded.
966.El
967.Pp
968.Fn bus_dmamap_unload
969will not perform any implicit synchronization of DMA buffers.
970This must be done explicitly by a call to
971.Fn bus_dmamap_sync
972prior to unloading the map.
973.It Fn bus_dmamap_sync "dmat" "map" "op"
974Performs synchronization of a device visible mapping with the CPU visible
975memory referenced by that mapping.
976Arguments are as follows:
977.Bl -tag -width ".Fa dmat"
978.It Fa dmat
979DMA tag used to allocate
980.Fa map .
981.It Fa map
982The DMA mapping to be synchronized.
983.It Fa op
984Type of synchronization operation to perform.
985See the definition of
986.Vt bus_dmasync_op_t
987for a description of the acceptable values for
988.Fa op .
989.El
990.Pp
991The
992.Fn bus_dmamap_sync
993function
994is the method used to ensure that CPU's and device's direct
995memory access (DMA) to shared
996memory is coherent.
997For example, the CPU might be used to set up the contents of a buffer
998that is to be made available to a device.
999To ensure that the data are visible via the device's mapping of that
1000memory, the buffer must be loaded and a DMA sync operation of
1001.Dv BUS_DMASYNC_PREWRITE
1002must be performed after the CPU has updated the buffer and before the device
1003access is initiated.
1004If the CPU modifies this buffer again later, another
1005.Dv BUS_DMASYNC_PREWRITE
1006sync operation must be performed before an additional device
1007access.
1008Conversely, suppose a device updates memory that is to be read by a CPU.
1009In this case, the buffer must be loaded, and a DMA sync operation of
1010.Dv BUS_DMASYNC_PREREAD
1011must be performed before the device access is initiated.
1012The CPU will only be able to see the results of this memory update
1013once the DMA operation has completed and a
1014.Dv BUS_DMASYNC_POSTREAD
1015sync operation has been performed.
1016.Pp
1017If read and write operations are not preceded and followed by the
1018appropriate synchronization operations, behavior is undefined.
1019.It Fn bus_dmamem_alloc "dmat" "**vaddr" "flags" "*mapp"
1020Allocates memory that is mapped into KVA at the address returned
1021in
1022.Fa vaddr
1023and that is permanently loaded into the newly created
1024.Vt bus_dmamap_t
1025returned via
1026.Fa mapp .
1027Arguments are as follows:
1028.Bl -tag -width ".Fa alignment"
1029.It Fa dmat
1030DMA tag describing the constraints of the DMA mapping.
1031.It Fa vaddr
1032Pointer to a pointer that will hold the returned KVA mapping of
1033the allocated region.
1034.It Fa flags
1035Flags are defined as follows:
1036.Bl -tag -width ".Dv BUS_DMA_NOWAIT"
1037.It Dv BUS_DMA_WAITOK
1038The routine can safely wait (sleep) for resources.
1039.It Dv BUS_DMA_NOWAIT
1040The routine is not allowed to wait for resources.
1041If resources are not available,
1042.Dv ENOMEM
1043is returned.
1044.It Dv BUS_DMA_COHERENT
1045Attempt to map this memory in a coherent fashion.
1046See
1047.Fn bus_dmamap_create
1048above for a description of this flag.
1049For
1050.Fn bus_dmamem_alloc ,
1051the
1052.Dv BUS_DMA_COHERENT
1053flag is currently implemented on arm, arm64 and sparc64.
1054.It Dv BUS_DMA_ZERO
1055Causes the allocated memory to be set to all zeros.
1056.It Dv BUS_DMA_NOCACHE
1057The allocated memory will not be cached in the processor caches.
1058All memory accesses appear on the bus and are executed
1059without reordering.
1060For
1061.Fn bus_dmamem_alloc ,
1062the
1063.Dv BUS_DMA_NOCACHE
1064flag is currently implemented on amd64 and i386 where it results in the
1065Strong Uncacheable PAT to be set for the allocated virtual address range.
1066.El
1067.It Fa mapp
1068Pointer to a
1069.Vt bus_dmamap_t
1070where the resulting DMA map will be stored.
1071.El
1072.Pp
1073The size of memory to be allocated is
1074.Fa maxsize
1075as specified in the call to
1076.Fn bus_dma_tag_create
1077for
1078.Fa dmat .
1079.Pp
1080The current implementation of
1081.Fn bus_dmamem_alloc
1082will allocate all requests as a single segment.
1083.Pp
1084An initial load operation is required to obtain the bus address of the allocated
1085memory, and an unload operation is required before freeing the memory, as
1086described below in
1087.Fn bus_dmamem_free .
1088Maps are automatically handled by this function and should not be explicitly
1089allocated or destroyed.
1090.Pp
1091Although an explicit load is not required for each access to the memory
1092referenced by the returned map, the synchronization requirements
1093as described in the
1094.Fn bus_dmamap_sync
1095section still apply and should be used to achieve portability on architectures
1096without coherent buses.
1097.Pp
1098Returns
1099.Er ENOMEM
1100if sufficient memory is not available for completing
1101the operation.
1102.It Fn bus_dmamem_free "dmat" "*vaddr" "map"
1103Frees memory previously allocated by
1104.Fn bus_dmamem_alloc .
1105Any mappings
1106will be invalidated.
1107Arguments are as follows:
1108.Bl -tag -width ".Fa vaddr"
1109.It Fa dmat
1110DMA tag.
1111.It Fa vaddr
1112Kernel virtual address of the memory.
1113.It Fa map
1114DMA map to be invalidated.
1115.El
1116.El
1117.Sh RETURN VALUES
1118Behavior is undefined if invalid arguments are passed to
1119any of the above functions.
1120If sufficient resources cannot be allocated for a given
1121transaction,
1122.Er ENOMEM
1123is returned.
1124All
1125routines that are not of type
1126.Vt void
1127will return 0 on success or an error
1128code on failure as discussed above.
1129.Pp
1130All
1131.Vt void
1132routines will succeed if provided with valid arguments.
1133.Sh LOCKING
1134Two locking protocols are used by
1135.Nm .
1136The first is a private global lock that is used to synchronize access to the
1137bounce buffer pool on the architectures that make use of them.
1138This lock is strictly a leaf lock that is only used internally to
1139.Nm
1140and is not exposed to clients of the API.
1141.Pp
1142The second protocol involves protecting various resources stored in the tag.
1143Since almost all
1144.Nm
1145operations are done through requests from the driver that created the tag,
1146the most efficient way to protect the tag resources is through the lock that
1147the driver uses.
1148In cases where
1149.Nm
1150acts on its own without being called by the driver, the lock primitive
1151specified in the tag is acquired and released automatically.
1152An example of this is when the
1153.Fn bus_dmamap_load
1154callback function is called from a deferred context instead of the driver
1155context.
1156This means that certain
1157.Nm
1158functions must always be called with the same lock held that is specified in the
1159tag.
1160These functions include:
1161.Pp
1162.Bl -item -offset indent -compact
1163.It
1164.Fn bus_dmamap_load
1165.It
1166.Fn bus_dmamap_load_bio
1167.It
1168.Fn bus_dmamap_load_ccb
1169.It
1170.Fn bus_dmamap_load_mbuf
1171.It
1172.Fn bus_dmamap_load_mbuf_sg
1173.It
1174.Fn bus_dmamap_load_uio
1175.It
1176.Fn bus_dmamap_unload
1177.It
1178.Fn bus_dmamap_sync
1179.El
1180.Pp
1181There is one exception to this rule.
1182It is common practice to call some of these functions during driver start-up
1183without any locks held.
1184So long as there is a guarantee of no possible concurrent use of the tag by
1185different threads during this operation, it is safe to not hold a lock for
1186these functions.
1187.Pp
1188Certain
1189.Nm
1190operations should not be called with the driver lock held, either because
1191they are already protected by an internal lock, or because they might sleep
1192due to memory or resource allocation.
1193The following functions must not be
1194called with any non-sleepable locks held:
1195.Pp
1196.Bl -item -offset indent -compact
1197.It
1198.Fn bus_dma_tag_create
1199.It
1200.Fn bus_dmamap_create
1201.It
1202.Fn bus_dmamem_alloc
1203.El
1204.Pp
1205All other functions do not have a locking protocol and can thus be
1206called with or without any system or driver locks held.
1207.Sh SEE ALSO
1208.Xr devclass 9 ,
1209.Xr device 9 ,
1210.Xr driver 9 ,
1211.Xr rman 9 ,
1212.Xr vslock 9
1213.Pp
1214.Rs
1215.%A "Jason R. Thorpe"
1216.%T "A Machine-Independent DMA Framework for NetBSD"
1217.%J "Proceedings of the Summer 1998 USENIX Technical Conference"
1218.%Q "USENIX Association"
1219.%D "June 1998"
1220.Re
1221.Sh HISTORY
1222The
1223.Nm
1224interface first appeared in
1225.Nx 1.3 .
1226.Pp
1227The
1228.Nm
1229API was adopted from
1230.Nx
1231for use in the CAM SCSI subsystem.
1232The alterations to the original API were aimed to remove the need for
1233a
1234.Vt bus_dma_segment_t
1235array stored in each
1236.Vt bus_dmamap_t
1237while allowing callers to queue up on scarce resources.
1238.Sh AUTHORS
1239The
1240.Nm
1241interface was designed and implemented by
1242.An Jason R. Thorpe
1243of the Numerical Aerospace Simulation Facility, NASA Ames Research Center.
1244Additional input on the
1245.Nm
1246design was provided by
1247.An -nosplit
1248.An Chris Demetriou ,
1249.An Charles Hannum ,
1250.An Ross Harvey ,
1251.An Matthew Jacob ,
1252.An Jonathan Stone ,
1253and
1254.An Matt Thomas .
1255.Pp
1256The
1257.Nm
1258interface in
1259.Fx
1260benefits from the contributions of
1261.An Justin T. Gibbs ,
1262.An Peter Wemm ,
1263.An Doug Rabson ,
1264.An Matthew N. Dodd ,
1265.An Sam Leffler ,
1266.An Maxime Henrion ,
1267.An Jake Burkholder ,
1268.An Takahashi Yoshihiro ,
1269.An Scott Long
1270and many others.
1271.Pp
1272This manual page was written by
1273.An Hiten M. Pandya
1274and
1275.An Justin T. Gibbs .
1276