xref: /freebsd/share/man/man9/bus_dma.9 (revision 65422c964ace1c1d0b7858e7f6f655ddeedb4db4)
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 May 25, 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_crp_buffer ,
73.Nm bus_dmamap_load_mbuf ,
74.Nm bus_dmamap_load_mbuf_sg ,
75.Nm bus_dmamap_load_uio ,
76.Nm bus_dmamap_unload ,
77.Nm bus_dmamap_sync ,
78.Nm bus_dmamem_alloc ,
79.Nm bus_dmamem_free
80.Nd Bus and Machine Independent DMA Mapping Interface
81.Sh SYNOPSIS
82.In machine/bus.h
83.Ft int
84.Fn bus_dma_tag_create "bus_dma_tag_t parent" "bus_size_t alignment" \
85"bus_addr_t boundary" "bus_addr_t lowaddr" "bus_addr_t highaddr" \
86"bus_dma_filter_t *filtfunc" "void *filtfuncarg" "bus_size_t maxsize" \
87"int nsegments" "bus_size_t maxsegsz" "int flags" "bus_dma_lock_t *lockfunc" \
88"void *lockfuncarg" "bus_dma_tag_t *dmat"
89.Ft int
90.Fn bus_dma_tag_destroy "bus_dma_tag_t dmat"
91.Ft void
92.Fo bus_dma_template_init
93.Fa "bus_dma_template_t template"
94.Fa "bus_dma_tag_t parent"
95.Fc
96.Ft int
97.Fo bus_dma_template_tag
98.Fa "bus_dma_template_t template"
99.Fa "bus_dma_tag_t *dmat"
100.Fc
101.Ft void
102.Fo bus_dma_template_clone
103.Fa "bus_dma_template_t template"
104.Fa "bus_dma_tag_t dmat"
105.Fc
106.Ft int
107.Fn bus_dmamap_create "bus_dma_tag_t dmat" "int flags" "bus_dmamap_t *mapp"
108.Ft int
109.Fn bus_dmamap_destroy "bus_dma_tag_t dmat" "bus_dmamap_t map"
110.Ft int
111.Fn bus_dmamap_load "bus_dma_tag_t dmat" "bus_dmamap_t map" "void *buf" \
112"bus_size_t buflen" "bus_dmamap_callback_t *callback" "void *callback_arg" \
113"int flags"
114.Ft int
115.Fn bus_dmamap_load_bio "bus_dma_tag_t dmat" "bus_dmamap_t map" \
116"struct bio *bio" "bus_dmamap_callback_t *callback" "void *callback_arg" \
117"int flags"
118.Ft int
119.Fn bus_dmamap_load_ccb "bus_dma_tag_t dmat" "bus_dmamap_t map" \
120"union ccb *ccb" "bus_dmamap_callback_t *callback" "void *callback_arg" \
121"int flags"
122.Ft int
123.Fn bus_dmamap_load_crp "bus_dma_tag_t dmat" "bus_dmamap_t map" \
124"struct crypto *crp" "bus_dmamap_callback_t *callback" "void *callback_arg" \
125"int flags"
126.Ft int
127.Fn bus_dmamap_load_crp_buffer "bus_dma_tag_t dmat" "bus_dmamap_t map" \
128"struct crypto_buffer *cb" "bus_dmamap_callback_t *callback" \
129"void *callback_arg" "int flags"
130.Ft int
131.Fn bus_dmamap_load_mbuf "bus_dma_tag_t dmat" "bus_dmamap_t map" \
132"struct mbuf *mbuf" "bus_dmamap_callback2_t *callback" "void *callback_arg" \
133"int flags"
134.Ft int
135.Fn bus_dmamap_load_mbuf_sg "bus_dma_tag_t dmat" "bus_dmamap_t map" \
136"struct mbuf *mbuf" "bus_dma_segment_t *segs" "int *nsegs" "int flags"
137.Ft int
138.Fn bus_dmamap_load_uio "bus_dma_tag_t dmat" "bus_dmamap_t map" \
139"struct uio *uio" "bus_dmamap_callback2_t *callback" "void *callback_arg" \
140"int flags"
141.Ft void
142.Fn bus_dmamap_unload "bus_dma_tag_t dmat" "bus_dmamap_t map"
143.Ft void
144.Fn bus_dmamap_sync "bus_dma_tag_t dmat" "bus_dmamap_t map" \
145"op"
146.Ft int
147.Fn bus_dmamem_alloc "bus_dma_tag_t dmat" "void **vaddr" \
148"int flags" "bus_dmamap_t *mapp"
149.Ft void
150.Fn bus_dmamem_free "bus_dma_tag_t dmat" "void *vaddr" \
151"bus_dmamap_t map"
152.Sh DESCRIPTION
153Direct Memory Access (DMA) is a method of transferring data
154without involving the CPU, thus providing higher performance.
155A DMA transaction can be achieved between device to memory,
156device to device, or memory to memory.
157.Pp
158The
159.Nm
160API is a bus, device, and machine-independent (MI) interface to
161DMA mechanisms.
162It provides the client with flexibility and simplicity by
163abstracting machine dependent issues like setting up
164DMA mappings, handling cache issues, bus specific features
165and limitations.
166.Sh OVERVIEW
167A tag structure
168.Vt ( bus_dma_tag_t )
169is used to describe the properties of a group of related DMA
170transactions.
171One way to view this is that a tag describes the limitations of a DMA engine.
172For example, if a DMA engine in a device is limited to 32-bit addresses,
173that limitation is specified by a parameter when creating the tag
174for that device.
175Similarly, a tag can be marked as requiring buffers whose addresses are
176aligned to a specific boundary.
177.Pp
178Some devices may require multiple tags to describe DMA
179transactions with differing properties.
180For example, a device might require 16-byte alignment of its descriptor ring
181while permitting arbitrary alignment of I/O buffers.
182In this case,
183the driver must create one tag for the descriptor ring and a separate tag for
184I/O buffers.
185If a device has restrictions that are common to all DMA transactions
186in addition to restrictions that differ between unrelated groups of
187transactions,
188the driver can first create a
189.Dq parent
190tag that decribes the common restrictions.
191The per-group tags can then inherit these restrictions from this
192.Dq parent
193tag rather than having to list them explicitly when creating the per-group tags.
194.Pp
195A mapping structure
196.Vt ( bus_dmamap_t )
197represents a mapping of a memory region for DMA.
198On systems with I/O MMUs,
199the mapping structure tracks any I/O MMU entries used by a request.
200For DMA requests that require bounce pages,
201the mapping tracks the bounce pages used.
202.Pp
203To prepare for one or more DMA transactions,
204a mapping must be bound to a memory region by calling one of the
205.Fn bus_dmamap_load
206functions.
207These functions configure the mapping which can include programming entries
208in an I/O MMU and/or allocating bounce pages.
209An output of these functions
210(either directly or indirectly by invoking a callback routine)
211is the list of scatter/gather address ranges a consumer can pass to a DMA
212engine to access the memory region.
213When a mapping is no longer needed,
214the mapping must be unloaded via
215.Fn bus_dmamap_unload .
216.Pp
217Before and after each DMA transaction,
218.Fn bus_dmamap_sync
219must be used to ensure that the correct data is used by the DMA engine and
220the CPU.
221If a mapping uses bounce pages,
222the sync operations copy data between the bounce pages and the memory region
223bound to the mapping.
224Sync operations also handle architecture-specific details such as CPU cache
225flushing and CPU memory operation ordering.
226.Sh STATIC VS DYNAMIC
227.Nm
228handles two types of DMA transactions: static and dynamic.
229Static transactions are used with a long-lived memory region that is reused
230for many transactions such as a descriptor ring.
231Dynamic transactions are used for transfers to or from transient buffers
232such as I/O buffers holding a network packet or disk block.
233Each transaction type uses a different subset of the
234.Nm
235API.
236.Ss Static Transactions
237Static transactions use memory regions allocated by
238.Nm .
239Each static memory region is allocated by calling
240.Fn bus_dmamem_alloc .
241This function requires a valid tag describing the properties of the
242DMA transactions to this region such as alignment or address restrictions.
243Multiple regions can share a single tag if they share the same restrictions.
244.Pp
245.Fn bus_dmamem_alloc
246allocates a memory region along with a mapping object.
247The associated tag, memory region, and mapping object must then be passed to
248.Fn bus_dmamap_load
249to bind the mapping to the allocated region and obtain the
250scatter/gather list.
251.Pp
252It is expected that
253.Fn bus_dmamem_alloc
254will attempt to allocate memory requiring less expensive sync operations
255(for example, implementations should not allocate regions requiring bounce
256pages),
257but sync operations should still be used.
258For example, a driver should use
259.Fn bus_dmamap_sync
260in an interrupt handler before reading descriptor ring entries written by the
261device prior to the interrupt.
262.Pp
263When a consumer is finished with a memory region,
264it should unload the mapping via
265.Fn bus_dmamap_unload
266and then release the memory region and mapping object via
267.Fn bus_dmamem_free .
268.Ss Dynamic Transactions
269Dynamic transactions map memory regions provided by other parts of the system.
270A tag must be created via
271.Fn bus_dma_tag_create
272to describe the DMA transactions to and from these memory regions,
273and a pool of mapping objects must be allocated via
274.Fn bus_dmamap_create
275to track the mappings of any in-flight transactions.
276.Pp
277When a consumer wishes to schedule a transaction for a memory region,
278the consumer must first obtain an unused mapping object from its pool
279of mapping objects.
280The memory region must be bound to the mapping object via one of the
281.Fn bus_dmamap_load
282functions.
283Before scheduling the transaction,
284the consumer should sync the memory region via
285.Fn bus_dmamap_sync
286with one or more of the
287.Dq PRE
288flags.
289After the transaction has completed,
290the consumer should sync the memory region via
291.Fn bus_dmamap_sync
292with one or more of the
293.Dq POST
294flags.
295The mapping can then be unloaded via
296.Fn bus_dmamap_unload ,
297and the mapping object can be returned to the pool of unused mapping objects.
298.Pp
299When a consumer is no longer scheduling DMA transactions,
300the mapping objects should be freed via
301.Fn bus_dmamap_destroy ,
302and the tag should be freed via
303.Fn bus_dma_tag_destroy .
304.Sh STRUCTURES AND TYPES
305.Bl -tag -width indent
306.It Vt bus_dma_tag_t
307A machine-dependent (MD) opaque type that describes the
308characteristics of a group of DMA transactions.
309DMA tags are organized into a hierarchy, with each child
310tag inheriting the restrictions of its parent.
311This allows all devices along the path of DMA transactions
312to contribute to the constraints of those transactions.
313.It Vt bus_dma_template_t
314A template structure for creating a
315.Fa bus_dma_tag_t
316from a set of defaults.
317Once initialized with
318.Fn bus_dma_template_init ,
319a driver can over-ride individual fields to suit its needs.
320The following fields have the indicated values:
321.Bd -literal
322	alignment	1
323	boundary	0
324	lowaddr		BUS_SPACE_MAXADDR
325	highaddr	BUS_SPACE_MAXADDR
326	maxsize		BUS_SPACE_MAXSIZE
327	nsegments	BUS_SPACE_UNRESTRICTED
328	maxsegsize	BUS_SPACE_MAXSIZE
329	flags		0
330	lockfunc	NULL
331	lockfuncarg	NULL
332.Ed
333.Pp
334Descriptions of each field are documented with
335.Fn bus_dma_tag_create .
336Note that the
337.Fa filtfunc
338and
339.Fa filtfuncarg
340attributes of the DMA tag are not supported with templates.
341.It Vt bus_dma_filter_t
342Client specified address filter having the format:
343.Bl -tag -width indent
344.It Ft int
345.Fn "client_filter" "void *filtarg" "bus_addr_t testaddr"
346.El
347.Pp
348Address filters can be specified during tag creation to allow
349for devices whose DMA address restrictions cannot be specified
350by a single window.
351The
352.Fa filtarg
353argument is specified by the client during tag creation to be passed to all
354invocations of the callback.
355The
356.Fa testaddr
357argument contains a potential starting address of a DMA mapping.
358The filter function operates on the set of addresses from
359.Fa testaddr
360to
361.Ql trunc_page(testaddr) + PAGE_SIZE - 1 ,
362inclusive.
363The filter function should return zero if any mapping in this range
364can be accommodated by the device and non-zero otherwise.
365.It Vt bus_dma_segment_t
366A machine-dependent type that describes individual
367DMA segments.
368It contains the following fields:
369.Bd -literal
370	bus_addr_t	ds_addr;
371	bus_size_t	ds_len;
372.Ed
373.Pp
374The
375.Fa ds_addr
376field contains the device visible address of the DMA segment, and
377.Fa ds_len
378contains the length of the DMA segment.
379Although the DMA segments returned by a mapping call will adhere to
380all restrictions necessary for a successful DMA operation, some conversion
381(e.g.\& a conversion from host byte order to the device's byte order) is
382almost always required when presenting segment information to the device.
383.It Vt bus_dmamap_t
384A machine-dependent opaque type describing an individual mapping.
385One map is used for each memory allocation that will be loaded.
386Maps can be reused once they have been unloaded.
387Multiple maps can be associated with one DMA tag.
388While the value of the map may evaluate to
389.Dv NULL
390on some platforms under certain conditions,
391it should never be assumed that it will be
392.Dv NULL
393in all cases.
394.It Vt bus_dmamap_callback_t
395Client specified callback for receiving mapping information resulting from
396the load of a
397.Vt bus_dmamap_t
398via
399.Fn bus_dmamap_load ,
400.Fn bus_dmamap_load_bio ,
401.Fn bus_dmamap_load_ccb ,
402.Fn bus_dmamap_load_crp ,
403or
404.Fn bus_dmamap_load_crp_buffer .
405Callbacks are of the format:
406.Bl -tag -width indent
407.It Ft void
408.Fn "client_callback" "void *callback_arg" "bus_dma_segment_t *segs" \
409"int nseg" "int error"
410.El
411.Pp
412The
413.Fa callback_arg
414is the callback argument passed to dmamap load functions.
415The
416.Fa segs
417and
418.Fa nseg
419arguments describe an array of
420.Vt bus_dma_segment_t
421structures that represent the mapping.
422This array is only valid within the scope of the callback function.
423The success or failure of the mapping is indicated by the
424.Fa error
425argument.
426More information on the use of callbacks can be found in the
427description of the individual dmamap load functions.
428.It Vt bus_dmamap_callback2_t
429Client specified callback for receiving mapping information resulting from
430the load of a
431.Vt bus_dmamap_t
432via
433.Fn bus_dmamap_load_uio
434or
435.Fn bus_dmamap_load_mbuf .
436.Pp
437Callback2s are of the format:
438.Bl -tag -width indent
439.It Ft void
440.Fn "client_callback2" "void *callback_arg" "bus_dma_segment_t *segs" \
441"int nseg" "bus_size_t mapsize" "int error"
442.El
443.Pp
444Callback2's behavior is the same as
445.Vt bus_dmamap_callback_t
446with the addition that the length of the data mapped is provided via
447.Fa mapsize .
448.It Vt bus_dmasync_op_t
449Memory synchronization operation specifier.
450Bus DMA requires explicit synchronization of memory with its device
451visible mapping in order to guarantee memory coherency.
452The
453.Vt bus_dmasync_op_t
454allows the type of DMA operation that will be or has been performed
455to be communicated to the system so that the correct coherency measures
456are taken.
457The operations are represented as bitfield flags that can be combined together,
458though it only makes sense to combine PRE flags or POST flags, not both.
459See the
460.Fn bus_dmamap_sync
461description below for more details on how to use these operations.
462.Pp
463All operations specified below are performed from the host memory point of view,
464where a read implies data coming from the device to the host memory, and a write
465implies data going from the host memory to the device.
466Alternatively, the operations can be thought of in terms of driver operations,
467where reading a network packet or storage sector corresponds to a read operation
468in
469.Nm .
470.Bl -tag -width ".Dv BUS_DMASYNC_POSTWRITE"
471.It Dv BUS_DMASYNC_PREREAD
472Perform any synchronization required prior to an update of host memory by the
473device.
474.It Dv BUS_DMASYNC_PREWRITE
475Perform any synchronization required after an update of host memory by the CPU
476and prior to device access to host memory.
477.It Dv BUS_DMASYNC_POSTREAD
478Perform any synchronization required after an update of host memory by the
479device and prior to CPU access to host memory.
480.It Dv BUS_DMASYNC_POSTWRITE
481Perform any synchronization required after device access to host memory.
482.El
483.It Vt bus_dma_lock_t
484Client specified lock/mutex manipulation method.
485This will be called from
486within busdma whenever a client lock needs to be manipulated.
487In its current form, the function will be called immediately before
488the callback for a DMA load operation that has been deferred with
489.Dv BUS_DMA_LOCK
490and immediately after with
491.Dv BUS_DMA_UNLOCK .
492If the load operation does not need to be deferred, then it
493will not be called since the function loading the map should
494be holding the appropriate locks.
495This method is of the format:
496.Bl -tag -width indent
497.It Ft void
498.Fn "lockfunc" "void *lockfunc_arg" "bus_dma_lock_op_t op"
499.El
500.Pp
501The
502.Fa lockfuncarg
503argument is specified by the client during tag creation to be passed to all
504invocations of the callback.
505The
506.Fa op
507argument specifies the lock operation to perform.
508.Pp
509Two
510.Vt lockfunc
511implementations are provided for convenience.
512.Fn busdma_lock_mutex
513performs standard mutex operations on the sleep mutex provided via
514.Fa lockfuncarg .
515.Fn dflt_lock
516will generate a system panic if it is called.
517It is substituted into the tag when
518.Fa lockfunc
519is passed as
520.Dv NULL
521to
522.Fn bus_dma_tag_create
523and is useful for tags that should not be used with deferred load operations.
524.It Vt bus_dma_lock_op_t
525Operations to be performed by the client-specified
526.Fn lockfunc .
527.Bl -tag -width ".Dv BUS_DMA_UNLOCK"
528.It Dv BUS_DMA_LOCK
529Acquires and/or locks the client locking primitive.
530.It Dv BUS_DMA_UNLOCK
531Releases and/or unlocks the client locking primitive.
532.El
533.El
534.Sh FUNCTIONS
535.Bl -tag -width indent
536.It Fn bus_dma_tag_create "parent" "alignment" "boundary" "lowaddr" \
537"highaddr" "*filtfunc" "*filtfuncarg" "maxsize" "nsegments" "maxsegsz" \
538"flags" "lockfunc" "lockfuncarg" "*dmat"
539Allocates a DMA tag, and initializes it according to
540the arguments provided:
541.Bl -tag -width ".Fa filtfuncarg"
542.It Fa parent
543A parent tag from which to inherit restrictions.
544The restrictions passed in other arguments can only further tighten the
545restrictions inherited from the parent tag.
546.Pp
547All tags created by a device driver must inherit from the tag returned by
548.Fn bus_get_dma_tag
549to honor restrictions between the parent bridge, CPU memory, and the
550device.
551.It Fa alignment
552Alignment constraint, in bytes, of any mappings created using this tag.
553The alignment must be a power of 2.
554Hardware that can DMA starting at any address would specify
555.Em 1
556for byte alignment.
557Hardware requiring DMA transfers to start on a multiple of 4K
558would specify
559.Em 4096 .
560.It Fa boundary
561Boundary constraint, in bytes, of the target DMA memory region.
562The boundary indicates the set of addresses, all multiples of the
563boundary argument, that cannot be crossed by a single
564.Vt bus_dma_segment_t .
565The boundary must be a power of 2 and must be no smaller than the
566maximum segment size.
567.Ql 0
568indicates that there are no boundary restrictions.
569.It Fa lowaddr , highaddr
570Bounds of the window of bus address space that
571.Em cannot
572be directly accessed by the device.
573The window contains all addresses greater than
574.Fa lowaddr
575and less than or equal to
576.Fa highaddr .
577For example, a device incapable of DMA above 4GB, would specify a
578.Fa highaddr
579of
580.Dv BUS_SPACE_MAXADDR
581and a
582.Fa lowaddr
583of
584.Dv BUS_SPACE_MAXADDR_32BIT .
585Similarly a device that can only perform DMA to addresses below
58616MB would specify a
587.Fa highaddr
588of
589.Dv BUS_SPACE_MAXADDR
590and a
591.Fa lowaddr
592of
593.Dv BUS_SPACE_MAXADDR_24BIT .
594Some implementations require that some region of device visible
595address space, overlapping available host memory, be outside the
596window.
597This area of
598.Ql safe memory
599is used to bounce requests that would otherwise conflict with
600the exclusion window.
601.It Fa filtfunc
602Optional filter function (may be
603.Dv NULL )
604to be called for any attempt to
605map memory into the window described by
606.Fa lowaddr
607and
608.Fa highaddr .
609A filter function is only required when the single window described
610by
611.Fa lowaddr
612and
613.Fa highaddr
614cannot adequately describe the constraints of the device.
615The filter function will be called for every machine page
616that overlaps the exclusion window.
617.It Fa filtfuncarg
618Argument passed to all calls to the filter function for this tag.
619May be
620.Dv NULL .
621.It Fa maxsize
622Maximum size, in bytes, of the sum of all segment lengths in a given
623DMA mapping associated with this tag.
624.It Fa nsegments
625Number of discontinuities (scatter/gather segments) allowed
626in a DMA mapped region.
627.It Fa maxsegsz
628Maximum size, in bytes, of a segment in any DMA mapped region associated
629with
630.Fa dmat .
631.It Fa flags
632Are as follows:
633.Bl -tag -width ".Dv BUS_DMA_ALLOCNOW"
634.It Dv BUS_DMA_ALLOCNOW
635Pre-allocate enough resources to handle at least one map load operation on
636this tag.
637If sufficient resources are not available,
638.Er ENOMEM
639is returned.
640This should not be used for tags that only describe buffers that will be
641allocated with
642.Fn bus_dmamem_alloc .
643Also, due to resource sharing with other tags, this flag does not guarantee
644that resources will be allocated or reserved exclusively for this tag.
645It should be treated only as a minor optimization.
646.It Dv BUS_DMA_COHERENT
647Indicate that the DMA engine and CPU are cache-coherent.
648Cached memory may be used to back allocations created by
649.Fn bus_dmamem_alloc .
650For
651.Fn bus_dma_tag_create ,
652the
653.Dv BUS_DMA_COHERENT
654flag is currently implemented on arm64.
655.El
656.It Fa lockfunc
657Optional lock manipulation function (may be
658.Dv NULL )
659to be called when busdma
660needs to manipulate a lock on behalf of the client.
661If
662.Dv NULL
663is specified,
664.Fn dflt_lock
665is used.
666.It Fa lockfuncarg
667Optional argument to be passed to the function specified by
668.Fa lockfunc .
669.It Fa dmat
670Pointer to a bus_dma_tag_t where the resulting DMA tag will
671be stored.
672.El
673.Pp
674Returns
675.Er ENOMEM
676if sufficient memory is not available for tag creation
677or allocating mapping resources.
678.It Fn bus_dma_tag_destroy "dmat"
679Deallocate the DMA tag
680.Fa dmat
681that was created by
682.Fn bus_dma_tag_create .
683.Pp
684Returns
685.Er EBUSY
686if any DMA maps remain associated with
687.Fa dmat
688or
689.Ql 0
690on success.
691.It Fn bus_dma_template_init "*template" "parent"
692Initializes a
693.Fa bus_dma_template_t
694structure and associates it with an optional
695.Fa parent .
696The
697.Fa parent
698argument may be NULL.
699.It Fn bus_dma_template_tag "*template" "*dmat"
700Unpacks a template into a tag, and returns the tag via the
701.Fa dmat .
702All return values are identical to
703.Fn bus_dma_tag_create .
704.It Fn bus_dma_template_clone "*template" "dmat"
705Clones the fields from a tag to a template.
706This is useful for cloning tags when paired with
707.Fn bus_dma_template_tag .
708A template that is filled in as a clone does not need to be initialized
709first.
710.It Fn bus_dmamap_create "dmat" "flags" "*mapp"
711Allocates and initializes a DMA map.
712Arguments are as follows:
713.Bl -tag -width ".Fa nsegments"
714.It Fa dmat
715DMA tag.
716.It Fa flags
717Are as follows:
718.Bl -tag -width ".Dv BUS_DMA_COHERENT"
719.It Dv BUS_DMA_COHERENT
720Attempt to map the memory loaded with this map such that cache sync
721operations are as cheap as possible.
722This flag is typically set on maps when the memory loaded with these will
723be accessed by both a CPU and a DMA engine, frequently such as control data
724and as opposed to streamable data such as receive and transmit buffers.
725Use of this flag does not remove the requirement of using
726.Fn bus_dmamap_sync ,
727but it may reduce the cost of performing these operations.
728.El
729.It Fa mapp
730Pointer to a
731.Vt bus_dmamap_t
732where the resulting DMA map will be stored.
733.El
734.Pp
735Returns
736.Er ENOMEM
737if sufficient memory is not available for creating the
738map or allocating mapping resources.
739.It Fn bus_dmamap_destroy "dmat" "map"
740Frees all resources associated with a given DMA map.
741Arguments are as follows:
742.Bl -tag -width ".Fa dmat"
743.It Fa dmat
744DMA tag used to allocate
745.Fa map .
746.It Fa map
747The DMA map to destroy.
748.El
749.Pp
750Returns
751.Er EBUSY
752if a mapping is still active for
753.Fa map .
754.It Fn bus_dmamap_load "dmat" "map" "buf" "buflen" "*callback" \
755"callback_arg" "flags"
756Creates a mapping in device visible address space of
757.Fa buflen
758bytes of
759.Fa buf ,
760associated with the DMA map
761.Fa map .
762This call will always return immediately and will not block for any reason.
763Arguments are as follows:
764.Bl -tag -width ".Fa buflen"
765.It Fa dmat
766DMA tag used to allocate
767.Fa map .
768.It Fa map
769A DMA map without a currently active mapping.
770.It Fa buf
771A kernel virtual address pointer to a contiguous (in KVA) buffer, to be
772mapped into device visible address space.
773.It Fa buflen
774The size of the buffer.
775.It Fa callback Fa callback_arg
776The callback function, and its argument.
777This function is called once sufficient mapping resources are available for
778the DMA operation.
779If resources are temporarily unavailable, this function will be deferred until
780later, but the load operation will still return immediately to the caller.
781Thus, callers should not assume that the callback will be called before the
782load returns, and code should be structured appropriately to handle this.
783See below for specific flags and error codes that control this behavior.
784.It Fa flags
785Are as follows:
786.Bl -tag -width ".Dv BUS_DMA_NOWAIT"
787.It Dv BUS_DMA_NOWAIT
788The load should not be deferred in case of insufficient mapping resources,
789and instead should return immediately with an appropriate error.
790.It Dv BUS_DMA_NOCACHE
791The generated transactions to and from the virtual page are non-cacheable.
792.El
793.El
794.Pp
795Return values to the caller are as follows:
796.Bl -tag -width ".Er EINPROGRESS"
797.It 0
798The callback has been called and completed.
799The status of the mapping has been delivered to the callback.
800.It Er EINPROGRESS
801The mapping has been deferred for lack of resources.
802The callback will be called as soon as resources are available.
803Callbacks are serviced in FIFO order.
804.Pp
805Note that subsequent load operations for the same tag that do not require
806extra resources will still succeed.
807This may result in out-of-order processing of requests.
808If the caller requires the order of requests to be preserved,
809then the caller is required to stall subsequent requests until a pending
810request's callback is invoked.
811.It Er ENOMEM
812The load request has failed due to insufficient resources, and the caller
813specifically used the
814.Dv BUS_DMA_NOWAIT
815flag.
816.It Er EINVAL
817The load request was invalid.
818The callback has been called and has been provided the same error.
819This error value may indicate that
820.Fa dmat ,
821.Fa map ,
822.Fa buf ,
823or
824.Fa callback
825were invalid, or
826.Fa buflen
827was larger than the
828.Fa maxsize
829argument used to create the dma tag
830.Fa dmat .
831.El
832.Pp
833When the callback is called, it is presented with an error value
834indicating the disposition of the mapping.
835Error may be one of the following:
836.Bl -tag -width ".Er EINPROGRESS"
837.It 0
838The mapping was successful and the
839.Fa dm_segs
840callback argument contains an array of
841.Vt bus_dma_segment_t
842elements describing the mapping.
843This array is only valid during the scope of the callback function.
844.It Er EFBIG
845A mapping could not be achieved within the segment constraints provided
846in the tag even though the requested allocation size was less than maxsize.
847.El
848.It Fn bus_dmamap_load_bio "dmat" "map" "bio" "callback" "callback_arg" "flags"
849This is a variation of
850.Fn bus_dmamap_load
851which maps buffers pointed to by
852.Fa bio
853for DMA transfers.
854.Fa bio
855may point to either a mapped or unmapped buffer.
856.It Fn bus_dmamap_load_ccb "dmat" "map" "ccb" "callback" "callback_arg" "flags"
857This is a variation of
858.Fn bus_dmamap_load
859which maps data pointed to by
860.Fa ccb
861for DMA transfers.
862The data for
863.Fa ccb
864may be any of the following types:
865.Bl -tag -width ".Er CAM_DATA_SG_PADDR"
866.It CAM_DATA_VADDR
867The data is a single KVA buffer.
868.It CAM_DATA_PADDR
869The data is a single bus address range.
870.It CAM_DATA_SG
871The data is a scatter/gather list of KVA buffers.
872.It CAM_DATA_SG_PADDR
873The data is a scatter/gather list of bus address ranges.
874.It CAM_DATA_BIO
875The data is contained in a
876.Vt struct bio
877attached to the CCB.
878.El
879.Pp
880.Fn bus_dmamap_load_ccb
881supports the following CCB XPT function codes:
882.Pp
883.Bl -item -offset indent -compact
884.It
885XPT_ATA_IO
886.It
887XPT_CONT_TARGET_IO
888.It
889XPT_SCSI_IO
890.El
891.It Fn bus_dmamap_load_crp "dmat" "map" "crp" "callback" "callback_arg" "flags"
892This is a variation of
893.Fn bus_dmamap_load
894which maps the input buffer pointed to by
895.Fa crp
896for DMA transfers.
897The
898.Dv BUS_DMA_NOWAIT
899flag is implied, thus no callback deferral will happen.
900.It Fn bus_dmamap_load_crp_buffer "dmat" "map" "cb" "callback" "callback_arg" \
901"flags"
902This is a variation of
903.Fn bus_dmamap_load
904which maps the crypto data buffer pointed to by
905.Fa cb
906for DMA transfers.
907The
908.Dv BUS_DMA_NOWAIT
909flag is implied, thus no callback deferral will happen.
910.It Fn bus_dmamap_load_mbuf "dmat" "map" "mbuf" "callback2" "callback_arg" \
911"flags"
912This is a variation of
913.Fn bus_dmamap_load
914which maps mbuf chains
915for DMA transfers.
916A
917.Vt bus_size_t
918argument is also passed to the callback routine, which
919contains the mbuf chain's packet header length.
920The
921.Dv BUS_DMA_NOWAIT
922flag is implied, thus no callback deferral will happen.
923.Pp
924Mbuf chains are assumed to be in kernel virtual address space.
925.Pp
926Beside the error values listed for
927.Fn bus_dmamap_load ,
928.Er EINVAL
929will be returned if the size of the mbuf chain exceeds the maximum limit of the
930DMA tag.
931.It Fn bus_dmamap_load_mbuf_sg "dmat" "map" "mbuf" "segs" "nsegs" "flags"
932This is just like
933.Fn bus_dmamap_load_mbuf
934except that it returns immediately without calling a callback function.
935It is provided for efficiency.
936The scatter/gather segment array
937.Va segs
938is provided by the caller and filled in directly by the function.
939The
940.Va nsegs
941argument is returned with the number of segments filled in.
942Returns the same errors as
943.Fn bus_dmamap_load_mbuf .
944.It Fn bus_dmamap_load_uio "dmat" "map" "uio" "callback2" "callback_arg" "flags"
945This is a variation of
946.Fn bus_dmamap_load
947which maps buffers pointed to by
948.Fa uio
949for DMA transfers.
950A
951.Vt bus_size_t
952argument is also passed to the callback routine, which contains the size of
953.Fa uio ,
954i.e.
955.Fa uio->uio_resid .
956The
957.Dv BUS_DMA_NOWAIT
958flag is implied, thus no callback deferral will happen.
959Returns the same errors as
960.Fn bus_dmamap_load .
961.Pp
962If
963.Fa uio->uio_segflg
964is
965.Dv UIO_USERSPACE ,
966then it is assumed that the buffer,
967.Fa uio
968is in
969.Fa "uio->uio_td->td_proc" Ns 's
970address space.
971User space memory must be in-core and wired prior to attempting a map
972load operation.
973Pages may be locked using
974.Xr vslock 9 .
975.It Fn bus_dmamap_unload "dmat" "map"
976Unloads a DMA map.
977Arguments are as follows:
978.Bl -tag -width ".Fa dmam"
979.It Fa dmat
980DMA tag used to allocate
981.Fa map .
982.It Fa map
983The DMA map that is to be unloaded.
984.El
985.Pp
986.Fn bus_dmamap_unload
987will not perform any implicit synchronization of DMA buffers.
988This must be done explicitly by a call to
989.Fn bus_dmamap_sync
990prior to unloading the map.
991.It Fn bus_dmamap_sync "dmat" "map" "op"
992Performs synchronization of a device visible mapping with the CPU visible
993memory referenced by that mapping.
994Arguments are as follows:
995.Bl -tag -width ".Fa dmat"
996.It Fa dmat
997DMA tag used to allocate
998.Fa map .
999.It Fa map
1000The DMA mapping to be synchronized.
1001.It Fa op
1002Type of synchronization operation to perform.
1003See the definition of
1004.Vt bus_dmasync_op_t
1005for a description of the acceptable values for
1006.Fa op .
1007.El
1008.Pp
1009The
1010.Fn bus_dmamap_sync
1011function
1012is the method used to ensure that CPU's and device's direct
1013memory access (DMA) to shared
1014memory is coherent.
1015For example, the CPU might be used to set up the contents of a buffer
1016that is to be made available to a device.
1017To ensure that the data are visible via the device's mapping of that
1018memory, the buffer must be loaded and a DMA sync operation of
1019.Dv BUS_DMASYNC_PREWRITE
1020must be performed after the CPU has updated the buffer and before the device
1021access is initiated.
1022If the CPU modifies this buffer again later, another
1023.Dv BUS_DMASYNC_PREWRITE
1024sync operation must be performed before an additional device
1025access.
1026Conversely, suppose a device updates memory that is to be read by a CPU.
1027In this case, the buffer must be loaded, and a DMA sync operation of
1028.Dv BUS_DMASYNC_PREREAD
1029must be performed before the device access is initiated.
1030The CPU will only be able to see the results of this memory update
1031once the DMA operation has completed and a
1032.Dv BUS_DMASYNC_POSTREAD
1033sync operation has been performed.
1034.Pp
1035If read and write operations are not preceded and followed by the
1036appropriate synchronization operations, behavior is undefined.
1037.It Fn bus_dmamem_alloc "dmat" "**vaddr" "flags" "*mapp"
1038Allocates memory that is mapped into KVA at the address returned
1039in
1040.Fa vaddr
1041and that is permanently loaded into the newly created
1042.Vt bus_dmamap_t
1043returned via
1044.Fa mapp .
1045Arguments are as follows:
1046.Bl -tag -width ".Fa alignment"
1047.It Fa dmat
1048DMA tag describing the constraints of the DMA mapping.
1049.It Fa vaddr
1050Pointer to a pointer that will hold the returned KVA mapping of
1051the allocated region.
1052.It Fa flags
1053Flags are defined as follows:
1054.Bl -tag -width ".Dv BUS_DMA_NOWAIT"
1055.It Dv BUS_DMA_WAITOK
1056The routine can safely wait (sleep) for resources.
1057.It Dv BUS_DMA_NOWAIT
1058The routine is not allowed to wait for resources.
1059If resources are not available,
1060.Dv ENOMEM
1061is returned.
1062.It Dv BUS_DMA_COHERENT
1063Attempt to map this memory in a coherent fashion.
1064See
1065.Fn bus_dmamap_create
1066above for a description of this flag.
1067For
1068.Fn bus_dmamem_alloc ,
1069the
1070.Dv BUS_DMA_COHERENT
1071flag is currently implemented on arm and arm64.
1072.It Dv BUS_DMA_ZERO
1073Causes the allocated memory to be set to all zeros.
1074.It Dv BUS_DMA_NOCACHE
1075The allocated memory will not be cached in the processor caches.
1076All memory accesses appear on the bus and are executed
1077without reordering.
1078For
1079.Fn bus_dmamem_alloc ,
1080the
1081.Dv BUS_DMA_NOCACHE
1082flag is currently implemented on amd64 and i386 where it results in the
1083Strong Uncacheable PAT to be set for the allocated virtual address range.
1084.El
1085.It Fa mapp
1086Pointer to a
1087.Vt bus_dmamap_t
1088where the resulting DMA map will be stored.
1089.El
1090.Pp
1091The size of memory to be allocated is
1092.Fa maxsize
1093as specified in the call to
1094.Fn bus_dma_tag_create
1095for
1096.Fa dmat .
1097.Pp
1098The current implementation of
1099.Fn bus_dmamem_alloc
1100will allocate all requests as a single segment.
1101.Pp
1102An initial load operation is required to obtain the bus address of the allocated
1103memory, and an unload operation is required before freeing the memory, as
1104described below in
1105.Fn bus_dmamem_free .
1106Maps are automatically handled by this function and should not be explicitly
1107allocated or destroyed.
1108.Pp
1109Although an explicit load is not required for each access to the memory
1110referenced by the returned map, the synchronization requirements
1111as described in the
1112.Fn bus_dmamap_sync
1113section still apply and should be used to achieve portability on architectures
1114without coherent buses.
1115.Pp
1116Returns
1117.Er ENOMEM
1118if sufficient memory is not available for completing
1119the operation.
1120.It Fn bus_dmamem_free "dmat" "*vaddr" "map"
1121Frees memory previously allocated by
1122.Fn bus_dmamem_alloc .
1123Any mappings
1124will be invalidated.
1125Arguments are as follows:
1126.Bl -tag -width ".Fa vaddr"
1127.It Fa dmat
1128DMA tag.
1129.It Fa vaddr
1130Kernel virtual address of the memory.
1131.It Fa map
1132DMA map to be invalidated.
1133.El
1134.El
1135.Sh RETURN VALUES
1136Behavior is undefined if invalid arguments are passed to
1137any of the above functions.
1138If sufficient resources cannot be allocated for a given
1139transaction,
1140.Er ENOMEM
1141is returned.
1142All
1143routines that are not of type
1144.Vt void
1145will return 0 on success or an error
1146code on failure as discussed above.
1147.Pp
1148All
1149.Vt void
1150routines will succeed if provided with valid arguments.
1151.Sh LOCKING
1152Two locking protocols are used by
1153.Nm .
1154The first is a private global lock that is used to synchronize access to the
1155bounce buffer pool on the architectures that make use of them.
1156This lock is strictly a leaf lock that is only used internally to
1157.Nm
1158and is not exposed to clients of the API.
1159.Pp
1160The second protocol involves protecting various resources stored in the tag.
1161Since almost all
1162.Nm
1163operations are done through requests from the driver that created the tag,
1164the most efficient way to protect the tag resources is through the lock that
1165the driver uses.
1166In cases where
1167.Nm
1168acts on its own without being called by the driver, the lock primitive
1169specified in the tag is acquired and released automatically.
1170An example of this is when the
1171.Fn bus_dmamap_load
1172callback function is called from a deferred context instead of the driver
1173context.
1174This means that certain
1175.Nm
1176functions must always be called with the same lock held that is specified in the
1177tag.
1178These functions include:
1179.Pp
1180.Bl -item -offset indent -compact
1181.It
1182.Fn bus_dmamap_load
1183.It
1184.Fn bus_dmamap_load_bio
1185.It
1186.Fn bus_dmamap_load_ccb
1187.It
1188.Fn bus_dmamap_load_mbuf
1189.It
1190.Fn bus_dmamap_load_mbuf_sg
1191.It
1192.Fn bus_dmamap_load_uio
1193.It
1194.Fn bus_dmamap_unload
1195.It
1196.Fn bus_dmamap_sync
1197.El
1198.Pp
1199There is one exception to this rule.
1200It is common practice to call some of these functions during driver start-up
1201without any locks held.
1202So long as there is a guarantee of no possible concurrent use of the tag by
1203different threads during this operation, it is safe to not hold a lock for
1204these functions.
1205.Pp
1206Certain
1207.Nm
1208operations should not be called with the driver lock held, either because
1209they are already protected by an internal lock, or because they might sleep
1210due to memory or resource allocation.
1211The following functions must not be
1212called with any non-sleepable locks held:
1213.Pp
1214.Bl -item -offset indent -compact
1215.It
1216.Fn bus_dma_tag_create
1217.It
1218.Fn bus_dmamap_create
1219.It
1220.Fn bus_dmamem_alloc
1221.El
1222.Pp
1223All other functions do not have a locking protocol and can thus be
1224called with or without any system or driver locks held.
1225.Sh SEE ALSO
1226.Xr devclass 9 ,
1227.Xr device 9 ,
1228.Xr driver 9 ,
1229.Xr rman 9 ,
1230.Xr vslock 9
1231.Pp
1232.Rs
1233.%A "Jason R. Thorpe"
1234.%T "A Machine-Independent DMA Framework for NetBSD"
1235.%J "Proceedings of the Summer 1998 USENIX Technical Conference"
1236.%Q "USENIX Association"
1237.%D "June 1998"
1238.Re
1239.Sh HISTORY
1240The
1241.Nm
1242interface first appeared in
1243.Nx 1.3 .
1244.Pp
1245The
1246.Nm
1247API was adopted from
1248.Nx
1249for use in the CAM SCSI subsystem.
1250The alterations to the original API were aimed to remove the need for
1251a
1252.Vt bus_dma_segment_t
1253array stored in each
1254.Vt bus_dmamap_t
1255while allowing callers to queue up on scarce resources.
1256.Sh AUTHORS
1257The
1258.Nm
1259interface was designed and implemented by
1260.An Jason R. Thorpe
1261of the Numerical Aerospace Simulation Facility, NASA Ames Research Center.
1262Additional input on the
1263.Nm
1264design was provided by
1265.An -nosplit
1266.An Chris Demetriou ,
1267.An Charles Hannum ,
1268.An Ross Harvey ,
1269.An Matthew Jacob ,
1270.An Jonathan Stone ,
1271and
1272.An Matt Thomas .
1273.Pp
1274The
1275.Nm
1276interface in
1277.Fx
1278benefits from the contributions of
1279.An Justin T. Gibbs ,
1280.An Peter Wemm ,
1281.An Doug Rabson ,
1282.An Matthew N. Dodd ,
1283.An Sam Leffler ,
1284.An Maxime Henrion ,
1285.An Jake Burkholder ,
1286.An Takahashi Yoshihiro ,
1287.An Scott Long
1288and many others.
1289.Pp
1290This manual page was written by
1291.An Hiten M. Pandya
1292and
1293.An Justin T. Gibbs .
1294