xref: /freebsd/share/man/man9/bus_dma.9 (revision c17dd0e88b5db94cf641e862f50bdad436704598)
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 April 14, 2020
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_crp ,
72.Nm bus_dmamap_load_mbuf ,
73.Nm bus_dmamap_load_mbuf_sg ,
74.Nm bus_dmamap_load_uio ,
75.Nm bus_dmamap_unload ,
76.Nm bus_dmamap_sync ,
77.Nm bus_dmamem_alloc ,
78.Nm bus_dmamem_free
79.Nd Bus and Machine Independent DMA Mapping Interface
80.Sh SYNOPSIS
81.In machine/bus.h
82.Ft int
83.Fn bus_dma_tag_create "bus_dma_tag_t parent" "bus_size_t alignment" \
84"bus_addr_t boundary" "bus_addr_t lowaddr" "bus_addr_t highaddr" \
85"bus_dma_filter_t *filtfunc" "void *filtfuncarg" "bus_size_t maxsize" \
86"int nsegments" "bus_size_t maxsegsz" "int flags" "bus_dma_lock_t *lockfunc" \
87"void *lockfuncarg" "bus_dma_tag_t *dmat"
88.Ft int
89.Fn bus_dma_tag_destroy "bus_dma_tag_t dmat"
90.Ft void
91.Fo bus_dma_template_init
92.Fa "bus_dma_template_t template"
93.Fa "bus_dma_tag_t parent"
94.Fc
95.Ft int
96.Fo bus_dma_template_tag
97.Fa "bus_dma_template_t template"
98.Fa "bus_dma_tag_t *dmat"
99.Fc
100.Ft void
101.Fo bus_dma_template_clone
102.Fa "bus_dma_template_t template"
103.Fa "bus_dma_tag_t dmat"
104.Fc
105.Ft int
106.Fn bus_dmamap_create "bus_dma_tag_t dmat" "int flags" "bus_dmamap_t *mapp"
107.Ft int
108.Fn bus_dmamap_destroy "bus_dma_tag_t dmat" "bus_dmamap_t map"
109.Ft int
110.Fn bus_dmamap_load "bus_dma_tag_t dmat" "bus_dmamap_t map" "void *buf" \
111"bus_size_t buflen" "bus_dmamap_callback_t *callback" "void *callback_arg" \
112"int flags"
113.Ft int
114.Fn bus_dmamap_load_bio "bus_dma_tag_t dmat" "bus_dmamap_t map" \
115"struct bio *bio" "bus_dmamap_callback_t *callback" "void *callback_arg" \
116"int flags"
117.Ft int
118.Fn bus_dmamap_load_ccb "bus_dma_tag_t dmat" "bus_dmamap_t map" \
119"union ccb *ccb" "bus_dmamap_callback_t *callback" "void *callback_arg" \
120"int flags"
121.Ft int
122.Fn bus_dmamap_load_crp "bus_dma_tag_t dmat" "bus_dmamap_t map" \
123"struct crypto *crp" "bus_dmamap_callback_t *callback" "void *callback_arg" \
124"int flags"
125.Ft int
126.Fn bus_dmamap_load_mbuf "bus_dma_tag_t dmat" "bus_dmamap_t map" \
127"struct mbuf *mbuf" "bus_dmamap_callback2_t *callback" "void *callback_arg" \
128"int flags"
129.Ft int
130.Fn bus_dmamap_load_mbuf_sg "bus_dma_tag_t dmat" "bus_dmamap_t map" \
131"struct mbuf *mbuf" "bus_dma_segment_t *segs" "int *nsegs" "int flags"
132.Ft int
133.Fn bus_dmamap_load_uio "bus_dma_tag_t dmat" "bus_dmamap_t map" \
134"struct uio *uio" "bus_dmamap_callback2_t *callback" "void *callback_arg" \
135"int flags"
136.Ft void
137.Fn bus_dmamap_unload "bus_dma_tag_t dmat" "bus_dmamap_t map"
138.Ft void
139.Fn bus_dmamap_sync "bus_dma_tag_t dmat" "bus_dmamap_t map" \
140"op"
141.Ft int
142.Fn bus_dmamem_alloc "bus_dma_tag_t dmat" "void **vaddr" \
143"int flags" "bus_dmamap_t *mapp"
144.Ft void
145.Fn bus_dmamem_free "bus_dma_tag_t dmat" "void *vaddr" \
146"bus_dmamap_t map"
147.Sh DESCRIPTION
148Direct Memory Access (DMA) is a method of transferring data
149without involving the CPU, thus providing higher performance.
150A DMA transaction can be achieved between device to memory,
151device to device, or memory to memory.
152.Pp
153The
154.Nm
155API is a bus, device, and machine-independent (MI) interface to
156DMA mechanisms.
157It provides the client with flexibility and simplicity by
158abstracting machine dependent issues like setting up
159DMA mappings, handling cache issues, bus specific features
160and limitations.
161.Sh OVERVIEW
162A tag structure
163.Vt ( bus_dma_tag_t )
164is used to describe the properties of a group of related DMA
165transactions.
166One way to view this is that a tag describes the limitations of a DMA engine.
167For example, if a DMA engine in a device is limited to 32-bit addresses,
168that limitation is specified by a parameter when creating the tag
169for that device.
170Similarly, a tag can be marked as requiring buffers whose addresses are
171aligned to a specific boundary.
172.Pp
173Some devices may require multiple tags to describe DMA
174transactions with differing properties.
175For example, a device might require 16-byte alignment of its descriptor ring
176while permitting arbitrary alignment of I/O buffers.
177In this case,
178the driver must create one tag for the descriptor ring and a separate tag for
179I/O buffers.
180If a device has restrictions that are common to all DMA transactions
181in addition to restrictions that differ between unrelated groups of
182transactions,
183the driver can first create a
184.Dq parent
185tag that decribes the common restrictions.
186The per-group tags can then inherit these restrictions from this
187.Dq parent
188tag rather than having to list them explicitly when creating the per-group tags.
189.Pp
190A mapping structure
191.Vt ( bus_dmamap_t )
192represents a mapping of a memory region for DMA.
193On systems with I/O MMUs,
194the mapping structure tracks any I/O MMU entries used by a request.
195For DMA requests that require bounce pages,
196the mapping tracks the bounce pages used.
197.Pp
198To prepare for one or more DMA transactions,
199a mapping must be bound to a memory region by calling one of the
200.Fn bus_dmamap_load
201functions.
202These functions configure the mapping which can include programming entries
203in an I/O MMU and/or allocating bounce pages.
204An output of these functions
205(either directly or indirectly by invoking a callback routine)
206is the list of scatter/gather address ranges a consumer can pass to a DMA
207engine to access the memory region.
208When a mapping is no longer needed,
209the mapping must be unloaded via
210.Fn bus_dmamap_unload .
211.Pp
212Before and after each DMA transaction,
213.Fn bus_dmamap_sync
214must be used to ensure that the correct data is used by the DMA engine and
215the CPU.
216If a mapping uses bounce pages,
217the sync operations copy data between the bounce pages and the memory region
218bound to the mapping.
219Sync operations also handle architecture-specific details such as CPU cache
220flushing and CPU memory operation ordering.
221.Sh STATIC VS DYNAMIC
222.Nm
223handles two types of DMA transactions: static and dynamic.
224Static transactions are used with a long-lived memory region that is reused
225for many transactions such as a descriptor ring.
226Dynamic transactions are used for transfers to or from transient buffers
227such as I/O buffers holding a network packet or disk block.
228Each transaction type uses a different subset of the
229.Nm
230API.
231.Ss Static Transactions
232Static transactions use memory regions allocated by
233.Nm .
234Each static memory region is allocated by calling
235.Fn bus_dmamem_alloc .
236This function requires a valid tag describing the properties of the
237DMA transactions to this region such as alignment or address restrictions.
238Multiple regions can share a single tag if they share the same restrictions.
239.Pp
240.Fn bus_dmamem_alloc
241allocates a memory region along with a mapping object.
242The associated tag, memory region, and mapping object must then be passed to
243.Fn bus_dmamap_load
244to bind the mapping to the allocated region and obtain the
245scatter/gather list.
246.Pp
247It is expected that
248.Fn bus_dmamem_alloc
249will attempt to allocate memory requiring less expensive sync operations
250(for example, implementations should not allocate regions requiring bounce
251pages),
252but sync operations should still be used.
253For example, a driver should use
254.Fn bus_dmamap_sync
255in an interrupt handler before reading descriptor ring entries written by the
256device prior to the interrupt.
257.Pp
258When a consumer is finished with a memory region,
259it should unload the mapping via
260.Fn bus_dmamap_unload
261and then release the memory region and mapping object via
262.Fn bus_dmamem_free .
263.Ss Dynamic Transactions
264Dynamic transactions map memory regions provided by other parts of the system.
265A tag must be created via
266.Fn bus_dma_tag_create
267to describe the DMA transactions to and from these memory regions,
268and a pool of mapping objects must be allocated via
269.Fn bus_dmamap_create
270to track the mappings of any in-flight transactions.
271.Pp
272When a consumer wishes to schedule a transaction for a memory region,
273the consumer must first obtain an unused mapping object from its pool
274of mapping objects.
275The memory region must be bound to the mapping object via one of the
276.Fn bus_dmamap_load
277functions.
278Before scheduling the transaction,
279the consumer should sync the memory region via
280.Fn bus_dmamap_sync
281with one or more of the
282.Dq PRE
283flags.
284After the transaction has completed,
285the consumer should sync the memory region via
286.Fn bus_dmamap_sync
287with one or more of the
288.Dq POST
289flags.
290The mapping can then be unloaded via
291.Fn bus_dmamap_unload ,
292and the mapping object can be returned to the pool of unused mapping objects.
293.Pp
294When a consumer is no longer scheduling DMA transactions,
295the mapping objects should be freed via
296.Fn bus_dmamap_destroy ,
297and the tag should be freed via
298.Fn bus_dma_tag_destroy .
299.Sh STRUCTURES AND TYPES
300.Bl -tag -width indent
301.It Vt bus_dma_tag_t
302A machine-dependent (MD) opaque type that describes the
303characteristics of a group of DMA transactions.
304DMA tags are organized into a hierarchy, with each child
305tag inheriting the restrictions of its parent.
306This allows all devices along the path of DMA transactions
307to contribute to the constraints of those transactions.
308.It Vt bus_dma_template_t
309A template structure for creating a
310.Fa bus_dma_tag_t
311from a set of defaults.
312Once initialized with
313.Fn bus_dma_template_init ,
314a driver can over-ride individual fields to suit its needs.
315The following fields have the indicated values:
316.Bd -literal
317	alignment	1
318	boundary	0
319	lowaddr		BUS_SPACE_MAXADDR
320	highaddr	BUS_SPACE_MAXADDR
321	maxsize		BUS_SPACE_MAXSIZE
322	nsegments	BUS_SPACE_UNRESTRICTED
323	maxsegsize	BUS_SPACE_MAXSIZE
324	flags		0
325	lockfunc	NULL
326	lockfuncarg	NULL
327.Ed
328.Pp
329Descriptions of each field are documented with
330.Fn bus_dma_tag_create .
331Note that the
332.Fa filtfunc
333and
334.Fa filtfuncarg
335attributes of the DMA tag are not supported with templates.
336.It Vt bus_dma_filter_t
337Client specified address filter having the format:
338.Bl -tag -width indent
339.It Ft int
340.Fn "client_filter" "void *filtarg" "bus_addr_t testaddr"
341.El
342.Pp
343Address filters can be specified during tag creation to allow
344for devices whose DMA address restrictions cannot be specified
345by a single window.
346The
347.Fa filtarg
348argument is specified by the client during tag creation to be passed to all
349invocations of the callback.
350The
351.Fa testaddr
352argument contains a potential starting address of a DMA mapping.
353The filter function operates on the set of addresses from
354.Fa testaddr
355to
356.Ql trunc_page(testaddr) + PAGE_SIZE - 1 ,
357inclusive.
358The filter function should return zero if any mapping in this range
359can be accommodated by the device and non-zero otherwise.
360.It Vt bus_dma_segment_t
361A machine-dependent type that describes individual
362DMA segments.
363It contains the following fields:
364.Bd -literal
365	bus_addr_t	ds_addr;
366	bus_size_t	ds_len;
367.Ed
368.Pp
369The
370.Fa ds_addr
371field contains the device visible address of the DMA segment, and
372.Fa ds_len
373contains the length of the DMA segment.
374Although the DMA segments returned by a mapping call will adhere to
375all restrictions necessary for a successful DMA operation, some conversion
376(e.g.\& a conversion from host byte order to the device's byte order) is
377almost always required when presenting segment information to the device.
378.It Vt bus_dmamap_t
379A machine-dependent opaque type describing an individual mapping.
380One map is used for each memory allocation that will be loaded.
381Maps can be reused once they have been unloaded.
382Multiple maps can be associated with one DMA tag.
383While the value of the map may evaluate to
384.Dv NULL
385on some platforms under certain conditions,
386it should never be assumed that it will be
387.Dv NULL
388in all cases.
389.It Vt bus_dmamap_callback_t
390Client specified callback for receiving mapping information resulting from
391the load of a
392.Vt bus_dmamap_t
393via
394.Fn bus_dmamap_load ,
395.Fn bus_dmamap_load_bio ,
396.Fn bus_dmamap_load_ccb ,
397or
398.Fn bus_dmamap_load_crp .
399Callbacks are of the format:
400.Bl -tag -width indent
401.It Ft void
402.Fn "client_callback" "void *callback_arg" "bus_dma_segment_t *segs" \
403"int nseg" "int error"
404.El
405.Pp
406The
407.Fa callback_arg
408is the callback argument passed to dmamap load functions.
409The
410.Fa segs
411and
412.Fa nseg
413arguments describe an array of
414.Vt bus_dma_segment_t
415structures that represent the mapping.
416This array is only valid within the scope of the callback function.
417The success or failure of the mapping is indicated by the
418.Fa error
419argument.
420More information on the use of callbacks can be found in the
421description of the individual dmamap load functions.
422.It Vt bus_dmamap_callback2_t
423Client specified callback for receiving mapping information resulting from
424the load of a
425.Vt bus_dmamap_t
426via
427.Fn bus_dmamap_load_uio
428or
429.Fn bus_dmamap_load_mbuf .
430.Pp
431Callback2s are of the format:
432.Bl -tag -width indent
433.It Ft void
434.Fn "client_callback2" "void *callback_arg" "bus_dma_segment_t *segs" \
435"int nseg" "bus_size_t mapsize" "int error"
436.El
437.Pp
438Callback2's behavior is the same as
439.Vt bus_dmamap_callback_t
440with the addition that the length of the data mapped is provided via
441.Fa mapsize .
442.It Vt bus_dmasync_op_t
443Memory synchronization operation specifier.
444Bus DMA requires explicit synchronization of memory with its device
445visible mapping in order to guarantee memory coherency.
446The
447.Vt bus_dmasync_op_t
448allows the type of DMA operation that will be or has been performed
449to be communicated to the system so that the correct coherency measures
450are taken.
451The operations are represented as bitfield flags that can be combined together,
452though it only makes sense to combine PRE flags or POST flags, not both.
453See the
454.Fn bus_dmamap_sync
455description below for more details on how to use these operations.
456.Pp
457All operations specified below are performed from the host memory point of view,
458where a read implies data coming from the device to the host memory, and a write
459implies data going from the host memory to the device.
460Alternatively, the operations can be thought of in terms of driver operations,
461where reading a network packet or storage sector corresponds to a read operation
462in
463.Nm .
464.Bl -tag -width ".Dv BUS_DMASYNC_POSTWRITE"
465.It Dv BUS_DMASYNC_PREREAD
466Perform any synchronization required prior to an update of host memory by the
467device.
468.It Dv BUS_DMASYNC_PREWRITE
469Perform any synchronization required after an update of host memory by the CPU
470and prior to device access to host memory.
471.It Dv BUS_DMASYNC_POSTREAD
472Perform any synchronization required after an update of host memory by the
473device and prior to CPU access to host memory.
474.It Dv BUS_DMASYNC_POSTWRITE
475Perform any synchronization required after device access to host memory.
476.El
477.It Vt bus_dma_lock_t
478Client specified lock/mutex manipulation method.
479This will be called from
480within busdma whenever a client lock needs to be manipulated.
481In its current form, the function will be called immediately before
482the callback for a DMA load operation that has been deferred with
483.Dv BUS_DMA_LOCK
484and immediately after with
485.Dv BUS_DMA_UNLOCK .
486If the load operation does not need to be deferred, then it
487will not be called since the function loading the map should
488be holding the appropriate locks.
489This method is of the format:
490.Bl -tag -width indent
491.It Ft void
492.Fn "lockfunc" "void *lockfunc_arg" "bus_dma_lock_op_t op"
493.El
494.Pp
495The
496.Fa lockfuncarg
497argument is specified by the client during tag creation to be passed to all
498invocations of the callback.
499The
500.Fa op
501argument specifies the lock operation to perform.
502.Pp
503Two
504.Vt lockfunc
505implementations are provided for convenience.
506.Fn busdma_lock_mutex
507performs standard mutex operations on the sleep mutex provided via
508.Fa lockfuncarg .
509.Fn dflt_lock
510will generate a system panic if it is called.
511It is substituted into the tag when
512.Fa lockfunc
513is passed as
514.Dv NULL
515to
516.Fn bus_dma_tag_create
517and is useful for tags that should not be used with deferred load operations.
518.It Vt bus_dma_lock_op_t
519Operations to be performed by the client-specified
520.Fn lockfunc .
521.Bl -tag -width ".Dv BUS_DMA_UNLOCK"
522.It Dv BUS_DMA_LOCK
523Acquires and/or locks the client locking primitive.
524.It Dv BUS_DMA_UNLOCK
525Releases and/or unlocks the client locking primitive.
526.El
527.El
528.Sh FUNCTIONS
529.Bl -tag -width indent
530.It Fn bus_dma_tag_create "parent" "alignment" "boundary" "lowaddr" \
531"highaddr" "*filtfunc" "*filtfuncarg" "maxsize" "nsegments" "maxsegsz" \
532"flags" "lockfunc" "lockfuncarg" "*dmat"
533Allocates a DMA tag, and initializes it according to
534the arguments provided:
535.Bl -tag -width ".Fa filtfuncarg"
536.It Fa parent
537A parent tag from which to inherit restrictions.
538The restrictions passed in other arguments can only further tighten the
539restrictions inherited from the parent tag.
540.Pp
541All tags created by a device driver must inherit from the tag returned by
542.Fn bus_get_dma_tag
543to honor restrictions between the parent bridge, CPU memory, and the
544device.
545.It Fa alignment
546Alignment constraint, in bytes, of any mappings created using this tag.
547The alignment must be a power of 2.
548Hardware that can DMA starting at any address would specify
549.Em 1
550for byte alignment.
551Hardware requiring DMA transfers to start on a multiple of 4K
552would specify
553.Em 4096 .
554.It Fa boundary
555Boundary constraint, in bytes, of the target DMA memory region.
556The boundary indicates the set of addresses, all multiples of the
557boundary argument, that cannot be crossed by a single
558.Vt bus_dma_segment_t .
559The boundary must be a power of 2 and must be no smaller than the
560maximum segment size.
561.Ql 0
562indicates that there are no boundary restrictions.
563.It Fa lowaddr , highaddr
564Bounds of the window of bus address space that
565.Em cannot
566be directly accessed by the device.
567The window contains all addresses greater than
568.Fa lowaddr
569and less than or equal to
570.Fa highaddr .
571For example, a device incapable of DMA above 4GB, would specify a
572.Fa highaddr
573of
574.Dv BUS_SPACE_MAXADDR
575and a
576.Fa lowaddr
577of
578.Dv BUS_SPACE_MAXADDR_32BIT .
579Similarly a device that can only perform DMA to addresses below
58016MB would specify a
581.Fa highaddr
582of
583.Dv BUS_SPACE_MAXADDR
584and a
585.Fa lowaddr
586of
587.Dv BUS_SPACE_MAXADDR_24BIT .
588Some implementations require that some region of device visible
589address space, overlapping available host memory, be outside the
590window.
591This area of
592.Ql safe memory
593is used to bounce requests that would otherwise conflict with
594the exclusion window.
595.It Fa filtfunc
596Optional filter function (may be
597.Dv NULL )
598to be called for any attempt to
599map memory into the window described by
600.Fa lowaddr
601and
602.Fa highaddr .
603A filter function is only required when the single window described
604by
605.Fa lowaddr
606and
607.Fa highaddr
608cannot adequately describe the constraints of the device.
609The filter function will be called for every machine page
610that overlaps the exclusion window.
611.It Fa filtfuncarg
612Argument passed to all calls to the filter function for this tag.
613May be
614.Dv NULL .
615.It Fa maxsize
616Maximum size, in bytes, of the sum of all segment lengths in a given
617DMA mapping associated with this tag.
618.It Fa nsegments
619Number of discontinuities (scatter/gather segments) allowed
620in a DMA mapped region.
621.It Fa maxsegsz
622Maximum size, in bytes, of a segment in any DMA mapped region associated
623with
624.Fa dmat .
625.It Fa flags
626Are as follows:
627.Bl -tag -width ".Dv BUS_DMA_ALLOCNOW"
628.It Dv BUS_DMA_ALLOCNOW
629Pre-allocate enough resources to handle at least one map load operation on
630this tag.
631If sufficient resources are not available,
632.Er ENOMEM
633is returned.
634This should not be used for tags that only describe buffers that will be
635allocated with
636.Fn bus_dmamem_alloc .
637Also, due to resource sharing with other tags, this flag does not guarantee
638that resources will be allocated or reserved exclusively for this tag.
639It should be treated only as a minor optimization.
640.It Dv BUS_DMA_COHERENT
641Indicate that the DMA engine and CPU are cache-coherent.
642Cached memory may be used to back allocations created by
643.Fn bus_dmamem_alloc .
644For
645.Fn bus_dma_tag_create ,
646the
647.Dv BUS_DMA_COHERENT
648flag is currently implemented on arm64.
649.El
650.It Fa lockfunc
651Optional lock manipulation function (may be
652.Dv NULL )
653to be called when busdma
654needs to manipulate a lock on behalf of the client.
655If
656.Dv NULL
657is specified,
658.Fn dflt_lock
659is used.
660.It Fa lockfuncarg
661Optional argument to be passed to the function specified by
662.Fa lockfunc .
663.It Fa dmat
664Pointer to a bus_dma_tag_t where the resulting DMA tag will
665be stored.
666.El
667.Pp
668Returns
669.Er ENOMEM
670if sufficient memory is not available for tag creation
671or allocating mapping resources.
672.It Fn bus_dma_tag_destroy "dmat"
673Deallocate the DMA tag
674.Fa dmat
675that was created by
676.Fn bus_dma_tag_create .
677.Pp
678Returns
679.Er EBUSY
680if any DMA maps remain associated with
681.Fa dmat
682or
683.Ql 0
684on success.
685.It Fn bus_dma_template_init "*template" "parent"
686Initializes a
687.Fa bus_dma_template_t
688structure and associates it with an optional
689.Fa parent .
690The
691.Fa parent
692argument may be NULL.
693.It Fn bus_dma_template_tag "*template" "*dmat"
694Unpacks a template into a tag, and returns the tag via the
695.Fa dmat .
696All return values are identical to
697.Fn bus_dma_tag_create .
698.It Fn bus_dma_template_clone "*template" "dmat"
699Clones the fields from a tag to a template.
700This is useful for cloning tags when paired with
701.Fn bus_dma_template_tag .
702A template that is filled in as a clone does not need to be initialized
703first.
704.It Fn bus_dmamap_create "dmat" "flags" "*mapp"
705Allocates and initializes a DMA map.
706Arguments are as follows:
707.Bl -tag -width ".Fa nsegments"
708.It Fa dmat
709DMA tag.
710.It Fa flags
711Are as follows:
712.Bl -tag -width ".Dv BUS_DMA_COHERENT"
713.It Dv BUS_DMA_COHERENT
714Attempt to map the memory loaded with this map such that cache sync
715operations are as cheap as possible.
716This flag is typically set on maps when the memory loaded with these will
717be accessed by both a CPU and a DMA engine, frequently such as control data
718and as opposed to streamable data such as receive and transmit buffers.
719Use of this flag does not remove the requirement of using
720.Fn bus_dmamap_sync ,
721but it may reduce the cost of performing these operations.
722.El
723.It Fa mapp
724Pointer to a
725.Vt bus_dmamap_t
726where the resulting DMA map will be stored.
727.El
728.Pp
729Returns
730.Er ENOMEM
731if sufficient memory is not available for creating the
732map or allocating mapping resources.
733.It Fn bus_dmamap_destroy "dmat" "map"
734Frees all resources associated with a given DMA map.
735Arguments are as follows:
736.Bl -tag -width ".Fa dmat"
737.It Fa dmat
738DMA tag used to allocate
739.Fa map .
740.It Fa map
741The DMA map to destroy.
742.El
743.Pp
744Returns
745.Er EBUSY
746if a mapping is still active for
747.Fa map .
748.It Fn bus_dmamap_load "dmat" "map" "buf" "buflen" "*callback" \
749"callback_arg" "flags"
750Creates a mapping in device visible address space of
751.Fa buflen
752bytes of
753.Fa buf ,
754associated with the DMA map
755.Fa map .
756This call will always return immediately and will not block for any reason.
757Arguments are as follows:
758.Bl -tag -width ".Fa buflen"
759.It Fa dmat
760DMA tag used to allocate
761.Fa map .
762.It Fa map
763A DMA map without a currently active mapping.
764.It Fa buf
765A kernel virtual address pointer to a contiguous (in KVA) buffer, to be
766mapped into device visible address space.
767.It Fa buflen
768The size of the buffer.
769.It Fa callback Fa callback_arg
770The callback function, and its argument.
771This function is called once sufficient mapping resources are available for
772the DMA operation.
773If resources are temporarily unavailable, this function will be deferred until
774later, but the load operation will still return immediately to the caller.
775Thus, callers should not assume that the callback will be called before the
776load returns, and code should be structured appropriately to handle this.
777See below for specific flags and error codes that control this behavior.
778.It Fa flags
779Are as follows:
780.Bl -tag -width ".Dv BUS_DMA_NOWAIT"
781.It Dv BUS_DMA_NOWAIT
782The load should not be deferred in case of insufficient mapping resources,
783and instead should return immediately with an appropriate error.
784.It Dv BUS_DMA_NOCACHE
785The generated transactions to and from the virtual page are non-cacheable.
786.El
787.El
788.Pp
789Return values to the caller are as follows:
790.Bl -tag -width ".Er EINPROGRESS"
791.It 0
792The callback has been called and completed.
793The status of the mapping has been delivered to the callback.
794.It Er EINPROGRESS
795The mapping has been deferred for lack of resources.
796The callback will be called as soon as resources are available.
797Callbacks are serviced in FIFO order.
798.Pp
799Note that subsequent load operations for the same tag that do not require
800extra resources will still succeed.
801This may result in out-of-order processing of requests.
802If the caller requires the order of requests to be preserved,
803then the caller is required to stall subsequent requests until a pending
804request's callback is invoked.
805.It Er ENOMEM
806The load request has failed due to insufficient resources, and the caller
807specifically used the
808.Dv BUS_DMA_NOWAIT
809flag.
810.It Er EINVAL
811The load request was invalid.
812The callback has been called and has been provided the same error.
813This error value may indicate that
814.Fa dmat ,
815.Fa map ,
816.Fa buf ,
817or
818.Fa callback
819were invalid, or
820.Fa buflen
821was larger than the
822.Fa maxsize
823argument used to create the dma tag
824.Fa dmat .
825.El
826.Pp
827When the callback is called, it is presented with an error value
828indicating the disposition of the mapping.
829Error may be one of the following:
830.Bl -tag -width ".Er EINPROGRESS"
831.It 0
832The mapping was successful and the
833.Fa dm_segs
834callback argument contains an array of
835.Vt bus_dma_segment_t
836elements describing the mapping.
837This array is only valid during the scope of the callback function.
838.It Er EFBIG
839A mapping could not be achieved within the segment constraints provided
840in the tag even though the requested allocation size was less than maxsize.
841.El
842.It Fn bus_dmamap_load_bio "dmat" "map" "bio" "callback" "callback_arg" "flags"
843This is a variation of
844.Fn bus_dmamap_load
845which maps buffers pointed to by
846.Fa bio
847for DMA transfers.
848.Fa bio
849may point to either a mapped or unmapped buffer.
850.It Fn bus_dmamap_load_ccb "dmat" "map" "ccb" "callback" "callback_arg" "flags"
851This is a variation of
852.Fn bus_dmamap_load
853which maps data pointed to by
854.Fa ccb
855for DMA transfers.
856The data for
857.Fa ccb
858may be any of the following types:
859.Bl -tag -width ".Er CAM_DATA_SG_PADDR"
860.It CAM_DATA_VADDR
861The data is a single KVA buffer.
862.It CAM_DATA_PADDR
863The data is a single bus address range.
864.It CAM_DATA_SG
865The data is a scatter/gather list of KVA buffers.
866.It CAM_DATA_SG_PADDR
867The data is a scatter/gather list of bus address ranges.
868.It CAM_DATA_BIO
869The data is contained in a
870.Vt struct bio
871attached to the CCB.
872.El
873.Pp
874.Fn bus_dmamap_load_ccb
875supports the following CCB XPT function codes:
876.Pp
877.Bl -item -offset indent -compact
878.It
879XPT_ATA_IO
880.It
881XPT_CONT_TARGET_IO
882.It
883XPT_SCSI_IO
884.El
885.It Fn bus_dmamap_load_crp "dmat" "map" "crp" "callback" "callback_arg" "flags"
886This is a variation of
887.Fn bus_dmamap_load
888which maps buffers pointed to by
889.Fa crp
890for DMA transfers.
891The
892.Dv BUS_DMA_NOWAIT
893flag is implied, thus no callback deferral will happen.
894.It Fn bus_dmamap_load_mbuf "dmat" "map" "mbuf" "callback2" "callback_arg" \
895"flags"
896This is a variation of
897.Fn bus_dmamap_load
898which maps mbuf chains
899for DMA transfers.
900A
901.Vt bus_size_t
902argument is also passed to the callback routine, which
903contains the mbuf chain's packet header length.
904The
905.Dv BUS_DMA_NOWAIT
906flag is implied, thus no callback deferral will happen.
907.Pp
908Mbuf chains are assumed to be in kernel virtual address space.
909.Pp
910Beside the error values listed for
911.Fn bus_dmamap_load ,
912.Er EINVAL
913will be returned if the size of the mbuf chain exceeds the maximum limit of the
914DMA tag.
915.It Fn bus_dmamap_load_mbuf_sg "dmat" "map" "mbuf" "segs" "nsegs" "flags"
916This is just like
917.Fn bus_dmamap_load_mbuf
918except that it returns immediately without calling a callback function.
919It is provided for efficiency.
920The scatter/gather segment array
921.Va segs
922is provided by the caller and filled in directly by the function.
923The
924.Va nsegs
925argument is returned with the number of segments filled in.
926Returns the same errors as
927.Fn bus_dmamap_load_mbuf .
928.It Fn bus_dmamap_load_uio "dmat" "map" "uio" "callback2" "callback_arg" "flags"
929This is a variation of
930.Fn bus_dmamap_load
931which maps buffers pointed to by
932.Fa uio
933for DMA transfers.
934A
935.Vt bus_size_t
936argument is also passed to the callback routine, which contains the size of
937.Fa uio ,
938i.e.
939.Fa uio->uio_resid .
940The
941.Dv BUS_DMA_NOWAIT
942flag is implied, thus no callback deferral will happen.
943Returns the same errors as
944.Fn bus_dmamap_load .
945.Pp
946If
947.Fa uio->uio_segflg
948is
949.Dv UIO_USERSPACE ,
950then it is assumed that the buffer,
951.Fa uio
952is in
953.Fa "uio->uio_td->td_proc" Ns 's
954address space.
955User space memory must be in-core and wired prior to attempting a map
956load operation.
957Pages may be locked using
958.Xr vslock 9 .
959.It Fn bus_dmamap_unload "dmat" "map"
960Unloads a DMA map.
961Arguments are as follows:
962.Bl -tag -width ".Fa dmam"
963.It Fa dmat
964DMA tag used to allocate
965.Fa map .
966.It Fa map
967The DMA map that is to be unloaded.
968.El
969.Pp
970.Fn bus_dmamap_unload
971will not perform any implicit synchronization of DMA buffers.
972This must be done explicitly by a call to
973.Fn bus_dmamap_sync
974prior to unloading the map.
975.It Fn bus_dmamap_sync "dmat" "map" "op"
976Performs synchronization of a device visible mapping with the CPU visible
977memory referenced by that mapping.
978Arguments are as follows:
979.Bl -tag -width ".Fa dmat"
980.It Fa dmat
981DMA tag used to allocate
982.Fa map .
983.It Fa map
984The DMA mapping to be synchronized.
985.It Fa op
986Type of synchronization operation to perform.
987See the definition of
988.Vt bus_dmasync_op_t
989for a description of the acceptable values for
990.Fa op .
991.El
992.Pp
993The
994.Fn bus_dmamap_sync
995function
996is the method used to ensure that CPU's and device's direct
997memory access (DMA) to shared
998memory is coherent.
999For example, the CPU might be used to set up the contents of a buffer
1000that is to be made available to a device.
1001To ensure that the data are visible via the device's mapping of that
1002memory, the buffer must be loaded and a DMA sync operation of
1003.Dv BUS_DMASYNC_PREWRITE
1004must be performed after the CPU has updated the buffer and before the device
1005access is initiated.
1006If the CPU modifies this buffer again later, another
1007.Dv BUS_DMASYNC_PREWRITE
1008sync operation must be performed before an additional device
1009access.
1010Conversely, suppose a device updates memory that is to be read by a CPU.
1011In this case, the buffer must be loaded, and a DMA sync operation of
1012.Dv BUS_DMASYNC_PREREAD
1013must be performed before the device access is initiated.
1014The CPU will only be able to see the results of this memory update
1015once the DMA operation has completed and a
1016.Dv BUS_DMASYNC_POSTREAD
1017sync operation has been performed.
1018.Pp
1019If read and write operations are not preceded and followed by the
1020appropriate synchronization operations, behavior is undefined.
1021.It Fn bus_dmamem_alloc "dmat" "**vaddr" "flags" "*mapp"
1022Allocates memory that is mapped into KVA at the address returned
1023in
1024.Fa vaddr
1025and that is permanently loaded into the newly created
1026.Vt bus_dmamap_t
1027returned via
1028.Fa mapp .
1029Arguments are as follows:
1030.Bl -tag -width ".Fa alignment"
1031.It Fa dmat
1032DMA tag describing the constraints of the DMA mapping.
1033.It Fa vaddr
1034Pointer to a pointer that will hold the returned KVA mapping of
1035the allocated region.
1036.It Fa flags
1037Flags are defined as follows:
1038.Bl -tag -width ".Dv BUS_DMA_NOWAIT"
1039.It Dv BUS_DMA_WAITOK
1040The routine can safely wait (sleep) for resources.
1041.It Dv BUS_DMA_NOWAIT
1042The routine is not allowed to wait for resources.
1043If resources are not available,
1044.Dv ENOMEM
1045is returned.
1046.It Dv BUS_DMA_COHERENT
1047Attempt to map this memory in a coherent fashion.
1048See
1049.Fn bus_dmamap_create
1050above for a description of this flag.
1051For
1052.Fn bus_dmamem_alloc ,
1053the
1054.Dv BUS_DMA_COHERENT
1055flag is currently implemented on arm and arm64.
1056.It Dv BUS_DMA_ZERO
1057Causes the allocated memory to be set to all zeros.
1058.It Dv BUS_DMA_NOCACHE
1059The allocated memory will not be cached in the processor caches.
1060All memory accesses appear on the bus and are executed
1061without reordering.
1062For
1063.Fn bus_dmamem_alloc ,
1064the
1065.Dv BUS_DMA_NOCACHE
1066flag is currently implemented on amd64 and i386 where it results in the
1067Strong Uncacheable PAT to be set for the allocated virtual address range.
1068.El
1069.It Fa mapp
1070Pointer to a
1071.Vt bus_dmamap_t
1072where the resulting DMA map will be stored.
1073.El
1074.Pp
1075The size of memory to be allocated is
1076.Fa maxsize
1077as specified in the call to
1078.Fn bus_dma_tag_create
1079for
1080.Fa dmat .
1081.Pp
1082The current implementation of
1083.Fn bus_dmamem_alloc
1084will allocate all requests as a single segment.
1085.Pp
1086An initial load operation is required to obtain the bus address of the allocated
1087memory, and an unload operation is required before freeing the memory, as
1088described below in
1089.Fn bus_dmamem_free .
1090Maps are automatically handled by this function and should not be explicitly
1091allocated or destroyed.
1092.Pp
1093Although an explicit load is not required for each access to the memory
1094referenced by the returned map, the synchronization requirements
1095as described in the
1096.Fn bus_dmamap_sync
1097section still apply and should be used to achieve portability on architectures
1098without coherent buses.
1099.Pp
1100Returns
1101.Er ENOMEM
1102if sufficient memory is not available for completing
1103the operation.
1104.It Fn bus_dmamem_free "dmat" "*vaddr" "map"
1105Frees memory previously allocated by
1106.Fn bus_dmamem_alloc .
1107Any mappings
1108will be invalidated.
1109Arguments are as follows:
1110.Bl -tag -width ".Fa vaddr"
1111.It Fa dmat
1112DMA tag.
1113.It Fa vaddr
1114Kernel virtual address of the memory.
1115.It Fa map
1116DMA map to be invalidated.
1117.El
1118.El
1119.Sh RETURN VALUES
1120Behavior is undefined if invalid arguments are passed to
1121any of the above functions.
1122If sufficient resources cannot be allocated for a given
1123transaction,
1124.Er ENOMEM
1125is returned.
1126All
1127routines that are not of type
1128.Vt void
1129will return 0 on success or an error
1130code on failure as discussed above.
1131.Pp
1132All
1133.Vt void
1134routines will succeed if provided with valid arguments.
1135.Sh LOCKING
1136Two locking protocols are used by
1137.Nm .
1138The first is a private global lock that is used to synchronize access to the
1139bounce buffer pool on the architectures that make use of them.
1140This lock is strictly a leaf lock that is only used internally to
1141.Nm
1142and is not exposed to clients of the API.
1143.Pp
1144The second protocol involves protecting various resources stored in the tag.
1145Since almost all
1146.Nm
1147operations are done through requests from the driver that created the tag,
1148the most efficient way to protect the tag resources is through the lock that
1149the driver uses.
1150In cases where
1151.Nm
1152acts on its own without being called by the driver, the lock primitive
1153specified in the tag is acquired and released automatically.
1154An example of this is when the
1155.Fn bus_dmamap_load
1156callback function is called from a deferred context instead of the driver
1157context.
1158This means that certain
1159.Nm
1160functions must always be called with the same lock held that is specified in the
1161tag.
1162These functions include:
1163.Pp
1164.Bl -item -offset indent -compact
1165.It
1166.Fn bus_dmamap_load
1167.It
1168.Fn bus_dmamap_load_bio
1169.It
1170.Fn bus_dmamap_load_ccb
1171.It
1172.Fn bus_dmamap_load_mbuf
1173.It
1174.Fn bus_dmamap_load_mbuf_sg
1175.It
1176.Fn bus_dmamap_load_uio
1177.It
1178.Fn bus_dmamap_unload
1179.It
1180.Fn bus_dmamap_sync
1181.El
1182.Pp
1183There is one exception to this rule.
1184It is common practice to call some of these functions during driver start-up
1185without any locks held.
1186So long as there is a guarantee of no possible concurrent use of the tag by
1187different threads during this operation, it is safe to not hold a lock for
1188these functions.
1189.Pp
1190Certain
1191.Nm
1192operations should not be called with the driver lock held, either because
1193they are already protected by an internal lock, or because they might sleep
1194due to memory or resource allocation.
1195The following functions must not be
1196called with any non-sleepable locks held:
1197.Pp
1198.Bl -item -offset indent -compact
1199.It
1200.Fn bus_dma_tag_create
1201.It
1202.Fn bus_dmamap_create
1203.It
1204.Fn bus_dmamem_alloc
1205.El
1206.Pp
1207All other functions do not have a locking protocol and can thus be
1208called with or without any system or driver locks held.
1209.Sh SEE ALSO
1210.Xr devclass 9 ,
1211.Xr device 9 ,
1212.Xr driver 9 ,
1213.Xr rman 9 ,
1214.Xr vslock 9
1215.Pp
1216.Rs
1217.%A "Jason R. Thorpe"
1218.%T "A Machine-Independent DMA Framework for NetBSD"
1219.%J "Proceedings of the Summer 1998 USENIX Technical Conference"
1220.%Q "USENIX Association"
1221.%D "June 1998"
1222.Re
1223.Sh HISTORY
1224The
1225.Nm
1226interface first appeared in
1227.Nx 1.3 .
1228.Pp
1229The
1230.Nm
1231API was adopted from
1232.Nx
1233for use in the CAM SCSI subsystem.
1234The alterations to the original API were aimed to remove the need for
1235a
1236.Vt bus_dma_segment_t
1237array stored in each
1238.Vt bus_dmamap_t
1239while allowing callers to queue up on scarce resources.
1240.Sh AUTHORS
1241The
1242.Nm
1243interface was designed and implemented by
1244.An Jason R. Thorpe
1245of the Numerical Aerospace Simulation Facility, NASA Ames Research Center.
1246Additional input on the
1247.Nm
1248design was provided by
1249.An -nosplit
1250.An Chris Demetriou ,
1251.An Charles Hannum ,
1252.An Ross Harvey ,
1253.An Matthew Jacob ,
1254.An Jonathan Stone ,
1255and
1256.An Matt Thomas .
1257.Pp
1258The
1259.Nm
1260interface in
1261.Fx
1262benefits from the contributions of
1263.An Justin T. Gibbs ,
1264.An Peter Wemm ,
1265.An Doug Rabson ,
1266.An Matthew N. Dodd ,
1267.An Sam Leffler ,
1268.An Maxime Henrion ,
1269.An Jake Burkholder ,
1270.An Takahashi Yoshihiro ,
1271.An Scott Long
1272and many others.
1273.Pp
1274This manual page was written by
1275.An Hiten M. Pandya
1276and
1277.An Justin T. Gibbs .
1278