Documentation/driver-api/dmaengine/provider.rst GitHub 원문 ↗

Linux 6.18.37 · Driver API / DMA Engine

DMAengine controller documentation

DMA controller hardware model, dma_device capability와 operation contract, metadata, descriptor reuse와 반복 transfer 설계를 설명합니다.

Source pathDocumentation/driver-api/dmaengine/provider.rst
Source versionLinux v6.18.37
TranslationDUJINLABS 전문 번역 + 해설

요약·해설과 원문, 전문 번역을 서로 분리했습니다. API 이름, symbol, source path는 원문 표기를 사용합니다.

1. 요약·해설

원문의 핵심 논리와 kernel programming 관점의 보충 설명입니다. 아래의 전문 번역과는 별도로 작성했습니다.

요약과 해설

provider.rst:1-659

DMAEngine provider는 channel과 DRQ request의 독립 관계, transfer width·burst·scatter-gather hardware model을 `dma_device` capability와 operation function으로 노출합니다. provider는 interrupt context에서의 prepare/status 제약, metadata buffer 소유권, residue granularity, asynchronous terminate 뒤 synchronize 보장을 정확히 구현해야 합니다. descriptor ACK/REUSE와 `DMA_PREP_REPEAT`·`DMA_PREP_LOAD_EOT` lifecycle을 지키고, 다음 transfer 선택을 interrupt handler에서 수행해 tasklet scheduling으로 생기는 channel idle latency를 줄이는 것이 핵심입니다.

2. 영어 원문 전체

번역 기준이 된 Linux v6.18.37 원문입니다. 줄 번호는 이 버전의 파일 좌표입니다.

원문 전체 펼치기
1 ==================================
2 DMAengine controller documentation
3 ==================================
4
5 Hardware Introduction
6 =====================
7
8 Most of the Slave DMA controllers have the same general principles of
9 operations.
10
11 They have a given number of channels to use for the DMA transfers, and
12 a given number of requests lines.
13
14 Requests and channels are pretty much orthogonal. Channels can be used
15 to serve several to any requests. To simplify, channels are the
16 entities that will be doing the copy, and requests what endpoints are
17 involved.
18
19 The request lines actually correspond to physical lines going from the
20 DMA-eligible devices to the controller itself. Whenever the device
21 will want to start a transfer, it will assert a DMA request (DRQ) by
22 asserting that request line.
23
24 A very simple DMA controller would only take into account a single
25 parameter: the transfer size. At each clock cycle, it would transfer a
26 byte of data from one buffer to another, until the transfer size has
27 been reached.
28
29 That wouldn't work well in the real world, since slave devices might
30 require a specific number of bits to be transferred in a single
31 cycle. For example, we may want to transfer as much data as the
32 physical bus allows to maximize performances when doing a simple
33 memory copy operation, but our audio device could have a narrower FIFO
34 that requires data to be written exactly 16 or 24 bits at a time. This
35 is why most if not all of the DMA controllers can adjust this, using a
36 parameter called the transfer width.
37
38 Moreover, some DMA controllers, whenever the RAM is used as a source
39 or destination, can group the reads or writes in memory into a buffer,
40 so instead of having a lot of small memory accesses, which is not
41 really efficient, you'll get several bigger transfers. This is done
42 using a parameter called the burst size, that defines how many single
43 reads/writes it's allowed to do without the controller splitting the
44 transfer into smaller sub-transfers.
45
46 Our theoretical DMA controller would then only be able to do transfers
47 that involve a single contiguous block of data. However, some of the
48 transfers we usually have are not, and want to copy data from
49 non-contiguous buffers to a contiguous buffer, which is called
50 scatter-gather.
51
52 DMAEngine, at least for mem2dev transfers, require support for
53 scatter-gather. So we're left with two cases here: either we have a
54 quite simple DMA controller that doesn't support it, and we'll have to
55 implement it in software, or we have a more advanced DMA controller,
56 that implements in hardware scatter-gather.
57
58 The latter are usually programmed using a collection of chunks to
59 transfer, and whenever the transfer is started, the controller will go
60 over that collection, doing whatever we programmed there.
61
62 This collection is usually either a table or a linked list. You will
63 then push either the address of the table and its number of elements,
64 or the first item of the list to one channel of the DMA controller,
65 and whenever a DRQ will be asserted, it will go through the collection
66 to know where to fetch the data from.
67
68 Either way, the format of this collection is completely dependent on
69 your hardware. Each DMA controller will require a different structure,
70 but all of them will require, for every chunk, at least the source and
71 destination addresses, whether it should increment these addresses or
72 not and the three parameters we saw earlier: the burst size, the
73 transfer width and the transfer size.
74
75 The one last thing is that usually, slave devices won't issue DRQ by
76 default, and you have to enable this in your slave device driver first
77 whenever you're willing to use DMA.
78
79 These were just the general memory-to-memory (also called mem2mem) or
80 memory-to-device (mem2dev) kind of transfers. Most devices often
81 support other kind of transfers or memory operations that dmaengine
82 support and will be detailed later in this document.
83
84 DMA Support in Linux
85 ====================
86
87 Historically, DMA controller drivers have been implemented using the
88 async TX API, to offload operations such as memory copy, XOR,
89 cryptography, etc., basically any memory to memory operation.
90
91 Over time, the need for memory to device transfers arose, and
92 dmaengine was extended. Nowadays, the async TX API is written as a
93 layer on top of dmaengine, and acts as a client. Still, dmaengine
94 accommodates that API in some cases, and made some design choices to
95 ensure that it stayed compatible.
96
97 For more information on the Async TX API, please look the relevant
98 documentation file in Documentation/crypto/async-tx-api.rst.
99
100 DMAEngine APIs
101 ==============
102
103 ``struct dma_device`` Initialization
104 ------------------------------------
105
106 Just like any other kernel framework, the whole DMAEngine registration
107 relies on the driver filling a structure and registering against the
108 framework. In our case, that structure is dma_device.
109
110 The first thing you need to do in your driver is to allocate this
111 structure. Any of the usual memory allocators will do, but you'll also
112 need to initialize a few fields in there:
113
114 - ``channels``: should be initialized as a list using the
115 INIT_LIST_HEAD macro for example
116
117 - ``src_addr_widths``:
118 should contain a bitmask of the supported source transfer width
119
120 - ``dst_addr_widths``:
121 should contain a bitmask of the supported destination transfer width
122
123 - ``directions``:
124 should contain a bitmask of the supported slave directions
125 (i.e. excluding mem2mem transfers)
126
127 - ``residue_granularity``:
128 granularity of the transfer residue reported to dma_set_residue.
129 This can be either:
130
131 - Descriptor:
132 your device doesn't support any kind of residue
133 reporting. The framework will only know that a particular
134 transaction descriptor is done.
135
136 - Segment:
137 your device is able to report which chunks have been transferred
138
139 - Burst:
140 your device is able to report which burst have been transferred
141
142 - ``dev``: should hold the pointer to the ``struct device`` associated
143 to your current driver instance.
144
145 Supported transaction types
146 ---------------------------
147
148 The next thing you need is to set which transaction types your device
149 (and driver) supports.
150
151 Our ``dma_device structure`` has a field called cap_mask that holds the
152 various types of transaction supported, and you need to modify this
153 mask using the dma_cap_set function, with various flags depending on
154 transaction types you support as an argument.
155
156 All those capabilities are defined in the ``dma_transaction_type enum``,
157 in ``include/linux/dmaengine.h``
158
159 Currently, the types available are:
160
161 - DMA_MEMCPY
162
163 - The device is able to do memory to memory copies
164
165 - No matter what the overall size of the combined chunks for source and
166 destination is, only as many bytes as the smallest of the two will be
167 transmitted. That means the number and size of the scatter-gather buffers in
168 both lists need not be the same, and that the operation functionally is
169 equivalent to a ``strncpy`` where the ``count`` argument equals the smallest
170 total size of the two scatter-gather list buffers.
171
172 - It's usually used for copying pixel data between host memory and
173 memory-mapped GPU device memory, such as found on modern PCI video graphics
174 cards. The most immediate example is the OpenGL API function
175 ``glReadPixels()``, which might require a verbatim copy of a huge
176 framebuffer from local device memory onto host memory.
177
178 - DMA_XOR
179
180 - The device is able to perform XOR operations on memory areas
181
182 - Used to accelerate XOR intensive tasks, such as RAID5
183
184 - DMA_XOR_VAL
185
186 - The device is able to perform parity check using the XOR
187 algorithm against a memory buffer.
188
189 - DMA_PQ
190
191 - The device is able to perform RAID6 P+Q computations, P being a
192 simple XOR, and Q being a Reed-Solomon algorithm.
193
194 - DMA_PQ_VAL
195
196 - The device is able to perform parity check using RAID6 P+Q
197 algorithm against a memory buffer.
198
199 - DMA_MEMSET
200
201 - The device is able to fill memory with the provided pattern
202
203 - The pattern is treated as a single byte signed value.
204
205 - DMA_INTERRUPT
206
207 - The device is able to trigger a dummy transfer that will
208 generate periodic interrupts
209
210 - Used by the client drivers to register a callback that will be
211 called on a regular basis through the DMA controller interrupt
212
213 - DMA_PRIVATE
214
215 - The devices only supports slave transfers, and as such isn't
216 available for async transfers.
217
218 - DMA_ASYNC_TX
219
220 - The device supports asynchronous memory-to-memory operations,
221 including memcpy, memset, xor, pq, xor_val, and pq_val.
222
223 - This capability is automatically set by the DMA engine
224 framework and must not be configured manually by device
225 drivers.
226
227 - DMA_SLAVE
228
229 - The device can handle device to memory transfers, including
230 scatter-gather transfers.
231
232 - While in the mem2mem case we were having two distinct types to
233 deal with a single chunk to copy or a collection of them, here,
234 we just have a single transaction type that is supposed to
235 handle both.
236
237 - If you want to transfer a single contiguous memory buffer,
238 simply build a scatter list with only one item.
239
240 - DMA_CYCLIC
241
242 - The device can handle cyclic transfers.
243
244 - A cyclic transfer is a transfer where the chunk collection will
245 loop over itself, with the last item pointing to the first.
246
247 - It's usually used for audio transfers, where you want to operate
248 on a single ring buffer that you will fill with your audio data.
249
250 - DMA_INTERLEAVE
251
252 - The device supports interleaved transfer.
253
254 - These transfers can transfer data from a non-contiguous buffer
255 to a non-contiguous buffer, opposed to DMA_SLAVE that can
256 transfer data from a non-contiguous data set to a continuous
257 destination buffer.
258
259 - It's usually used for 2d content transfers, in which case you
260 want to transfer a portion of uncompressed data directly to the
261 display to print it
262
263 - DMA_COMPLETION_NO_ORDER
264
265 - The device does not support in order completion.
266
267 - The driver should return DMA_OUT_OF_ORDER for device_tx_status if
268 the device is setting this capability.
269
270 - All cookie tracking and checking API should be treated as invalid if
271 the device exports this capability.
272
273 - At this point, this is incompatible with polling option for dmatest.
274
275 - If this cap is set, the user is recommended to provide an unique
276 identifier for each descriptor sent to the DMA device in order to
277 properly track the completion.
278
279 - DMA_REPEAT
280
281 - The device supports repeated transfers. A repeated transfer, indicated by
282 the DMA_PREP_REPEAT transfer flag, is similar to a cyclic transfer in that
283 it gets automatically repeated when it ends, but can additionally be
284 replaced by the client.
285
286 - This feature is limited to interleaved transfers, this flag should thus not
287 be set if the DMA_INTERLEAVE flag isn't set. This limitation is based on
288 the current needs of DMA clients, support for additional transfer types
289 should be added in the future if and when the need arises.
290
291 - DMA_LOAD_EOT
292
293 - The device supports replacing repeated transfers at end of transfer (EOT)
294 by queuing a new transfer with the DMA_PREP_LOAD_EOT flag set.
295
296 - Support for replacing a currently running transfer at another point (such
297 as end of burst instead of end of transfer) will be added in the future
298 based on DMA clients needs, if and when the need arises.
299
300 These various types will also affect how the source and destination
301 addresses change over time.
302
303 Addresses pointing to RAM are typically incremented (or decremented)
304 after each transfer. In case of a ring buffer, they may loop
305 (DMA_CYCLIC). Addresses pointing to a device's register (e.g. a FIFO)
306 are typically fixed.
307
308 Per descriptor metadata support
309 -------------------------------
310 Some data movement architecture (DMA controller and peripherals) uses metadata
311 associated with a transaction. The DMA controller role is to transfer the
312 payload and the metadata alongside.
313 The metadata itself is not used by the DMA engine itself, but it contains
314 parameters, keys, vectors, etc for peripheral or from the peripheral.
315
316 The DMAengine framework provides a generic ways to facilitate the metadata for
317 descriptors. Depending on the architecture the DMA driver can implement either
318 or both of the methods and it is up to the client driver to choose which one
319 to use.
320
321 - DESC_METADATA_CLIENT
322
323 The metadata buffer is allocated/provided by the client driver and it is
324 attached (via the dmaengine_desc_attach_metadata() helper to the descriptor.
325
326 From the DMA driver the following is expected for this mode:
327
328 - DMA_MEM_TO_DEV / DEV_MEM_TO_MEM
329
330 The data from the provided metadata buffer should be prepared for the DMA
331 controller to be sent alongside of the payload data. Either by copying to a
332 hardware descriptor, or highly coupled packet.
333
334 - DMA_DEV_TO_MEM
335
336 On transfer completion the DMA driver must copy the metadata to the client
337 provided metadata buffer before notifying the client about the completion.
338 After the transfer completion, DMA drivers must not touch the metadata
339 buffer provided by the client.
340
341 - DESC_METADATA_ENGINE
342
343 The metadata buffer is allocated/managed by the DMA driver. The client driver
344 can ask for the pointer, maximum size and the currently used size of the
345 metadata and can directly update or read it. dmaengine_desc_get_metadata_ptr()
346 and dmaengine_desc_set_metadata_len() is provided as helper functions.
347
348 From the DMA driver the following is expected for this mode:
349
350 - get_metadata_ptr()
351
352 Should return a pointer for the metadata buffer, the maximum size of the
353 metadata buffer and the currently used / valid (if any) bytes in the buffer.
354
355 - set_metadata_len()
356
357 It is called by the clients after it have placed the metadata to the buffer
358 to let the DMA driver know the number of valid bytes provided.
359
360 Note: since the client will ask for the metadata pointer in the completion
361 callback (in DMA_DEV_TO_MEM case) the DMA driver must ensure that the
362 descriptor is not freed up prior the callback is called.
363
364 Device operations
365 -----------------
366
367 Our dma_device structure also requires a few function pointers in
368 order to implement the actual logic, now that we described what
369 operations we were able to perform.
370
371 The functions that we have to fill in there, and hence have to
372 implement, obviously depend on the transaction types you reported as
373 supported.
374
375 - ``device_alloc_chan_resources``
376
377 - ``device_free_chan_resources``
378
379 - These functions will be called whenever a driver will call
380 ``dma_request_channel`` or ``dma_release_channel`` for the first/last
381 time on the channel associated to that driver.
382
383 - They are in charge of allocating/freeing all the needed
384 resources in order for that channel to be useful for your driver.
385
386 - These functions can sleep.
387
388 - ``device_prep_dma_*``
389
390 - These functions are matching the capabilities you registered
391 previously.
392
393 - These functions all take the buffer or the scatterlist relevant
394 for the transfer being prepared, and should create a hardware
395 descriptor or a list of hardware descriptors from it
396
397 - These functions can be called from an interrupt context
398
399 - Any allocation you might do should be using the GFP_NOWAIT
400 flag, in order not to potentially sleep, but without depleting
401 the emergency pool either.
402
403 - Drivers should try to pre-allocate any memory they might need
404 during the transfer setup at probe time to avoid putting to
405 much pressure on the nowait allocator.
406
407 - It should return a unique instance of the
408 ``dma_async_tx_descriptor structure``, that further represents this
409 particular transfer.
410
411 - This structure can be initialized using the function
412 ``dma_async_tx_descriptor_init``.
413
414 - You'll also need to set two fields in this structure:
415
416 - flags:
417 TODO: Can it be modified by the driver itself, or
418 should it be always the flags passed in the arguments
419
420 - tx_submit: A pointer to a function you have to implement,
421 that is supposed to push the current transaction descriptor to a
422 pending queue, waiting for issue_pending to be called.
423
424 - In this structure the function pointer callback_result can be
425 initialized in order for the submitter to be notified that a
426 transaction has completed. In the earlier code the function pointer
427 callback has been used. However it does not provide any status to the
428 transaction and will be deprecated. The result structure defined as
429 ``dmaengine_result`` that is passed in to callback_result
430 has two fields:
431
432 - result: This provides the transfer result defined by
433 ``dmaengine_tx_result``. Either success or some error condition.
434
435 - residue: Provides the residue bytes of the transfer for those that
436 support residue.
437
438 - ``device_prep_peripheral_dma_vec``
439
440 - Similar to ``device_prep_slave_sg``, but it takes a pointer to a
441 array of ``dma_vec`` structures, which (in the long run) will replace
442 scatterlists.
443
444 - ``device_issue_pending``
445
446 - Takes the first transaction descriptor in the pending queue,
447 and starts the transfer. Whenever that transfer is done, it
448 should move to the next transaction in the list.
449
450 - This function can be called in an interrupt context
451
452 - ``device_tx_status``
453
454 - Should report the bytes left to go over on the given channel
455
456 - Should only care about the transaction descriptor passed as
457 argument, not the currently active one on a given channel
458
459 - The tx_state argument might be NULL
460
461 - Should use dma_set_residue to report it
462
463 - In the case of a cyclic transfer, it should only take into
464 account the total size of the cyclic buffer.
465
466 - Should return DMA_OUT_OF_ORDER if the device does not support in order
467 completion and is completing the operation out of order.
468
469 - This function can be called in an interrupt context.
470
471 - device_config
472
473 - Reconfigures the channel with the configuration given as argument
474
475 - This command should NOT perform synchronously, or on any
476 currently queued transfers, but only on subsequent ones
477
478 - In this case, the function will receive a ``dma_slave_config``
479 structure pointer as an argument, that will detail which
480 configuration to use.
481
482 - Even though that structure contains a direction field, this
483 field is deprecated in favor of the direction argument given to
484 the prep_* functions
485
486 - This call is mandatory for slave operations only. This should NOT be
487 set or expected to be set for memcpy operations.
488 If a driver support both, it should use this call for slave
489 operations only and not for memcpy ones.
490
491 - device_pause
492
493 - Pauses a transfer on the channel
494
495 - This command should operate synchronously on the channel,
496 pausing right away the work of the given channel
497
498 - device_resume
499
500 - Resumes a transfer on the channel
501
502 - This command should operate synchronously on the channel,
503 resuming right away the work of the given channel
504
505 - device_terminate_all
506
507 - Aborts all the pending and ongoing transfers on the channel
508
509 - For aborted transfers the complete callback should not be called
510
511 - Can be called from atomic context or from within a complete
512 callback of a descriptor. Must not sleep. Drivers must be able
513 to handle this correctly.
514
515 - Termination may be asynchronous. The driver does not have to
516 wait until the currently active transfer has completely stopped.
517 See device_synchronize.
518
519 - device_synchronize
520
521 - Must synchronize the termination of a channel to the current
522 context.
523
524 - Must make sure that memory for previously submitted
525 descriptors is no longer accessed by the DMA controller.
526
527 - Must make sure that all complete callbacks for previously
528 submitted descriptors have finished running and none are
529 scheduled to run.
530
531 - May sleep.
532
533
534 Misc notes
535 ==========
536
537 (stuff that should be documented, but don't really know
538 where to put them)
539
540 ``dma_run_dependencies``
541
542 - Should be called at the end of an async TX transfer, and can be
543 ignored in the slave transfers case.
544
545 - Makes sure that dependent operations are run before marking it
546 as complete.
547
548 dma_cookie_t
549
550 - it's a DMA transaction ID that will increment over time.
551
552 - Not really relevant any more since the introduction of ``virt-dma``
553 that abstracts it away.
554
555 dma_vec
556
557 - A small structure that contains a DMA address and length.
558
559 DMA_CTRL_ACK
560
561 - If clear, the descriptor cannot be reused by provider until the
562 client acknowledges receipt, i.e. has a chance to establish any
563 dependency chains
564
565 - This can be acked by invoking async_tx_ack()
566
567 - If set, does not mean descriptor can be reused
568
569 DMA_CTRL_REUSE
570
571 - If set, the descriptor can be reused after being completed. It should
572 not be freed by provider if this flag is set.
573
574 - The descriptor should be prepared for reuse by invoking
575 ``dmaengine_desc_set_reuse()`` which will set DMA_CTRL_REUSE.
576
577 - ``dmaengine_desc_set_reuse()`` will succeed only when channel support
578 reusable descriptor as exhibited by capabilities
579
580 - As a consequence, if a device driver wants to skip the
581 ``dma_map_sg()`` and ``dma_unmap_sg()`` in between 2 transfers,
582 because the DMA'd data wasn't used, it can resubmit the transfer right after
583 its completion.
584
585 - Descriptor can be freed in few ways
586
587 - Clearing DMA_CTRL_REUSE by invoking
588 ``dmaengine_desc_clear_reuse()`` and submitting for last txn
589
590 - Explicitly invoking ``dmaengine_desc_free()``, this can succeed only
591 when DMA_CTRL_REUSE is already set
592
593 - Terminating the channel
594
595 - DMA_PREP_CMD
596
597 - If set, the client driver tells DMA controller that passed data in DMA
598 API is command data.
599
600 - Interpretation of command data is DMA controller specific. It can be
601 used for issuing commands to other peripherals/register reads/register
602 writes for which the descriptor should be in different format from
603 normal data descriptors.
604
605 - DMA_PREP_REPEAT
606
607 - If set, the transfer will be automatically repeated when it ends until a
608 new transfer is queued on the same channel with the DMA_PREP_LOAD_EOT flag.
609 If the next transfer to be queued on the channel does not have the
610 DMA_PREP_LOAD_EOT flag set, the current transfer will be repeated until the
611 client terminates all transfers.
612
613 - This flag is only supported if the channel reports the DMA_REPEAT
614 capability.
615
616 - DMA_PREP_LOAD_EOT
617
618 - If set, the transfer will replace the transfer currently being executed at
619 the end of the transfer.
620
621 - This is the default behaviour for non-repeated transfers, specifying
622 DMA_PREP_LOAD_EOT for non-repeated transfers will thus make no difference.
623
624 - When using repeated transfers, DMA clients will usually need to set the
625 DMA_PREP_LOAD_EOT flag on all transfers, otherwise the channel will keep
626 repeating the last repeated transfer and ignore the new transfers being
627 queued. Failure to set DMA_PREP_LOAD_EOT will appear as if the channel was
628 stuck on the previous transfer.
629
630 - This flag is only supported if the channel reports the DMA_LOAD_EOT
631 capability.
632
633 General Design Notes
634 ====================
635
636 Most of the DMAEngine drivers you'll see are based on a similar design
637 that handles the end of transfer interrupts in the handler, but defer
638 most work to a tasklet, including the start of a new transfer whenever
639 the previous transfer ended.
640
641 This is a rather inefficient design though, because the inter-transfer
642 latency will be not only the interrupt latency, but also the
643 scheduling latency of the tasklet, which will leave the channel idle
644 in between, which will slow down the global transfer rate.
645
646 You should avoid this kind of practice, and instead of electing a new
647 transfer in your tasklet, move that part to the interrupt handler in
648 order to have a shorter idle window (that we can't really avoid
649 anyway).
650
651 Glossary
652 ========
653
654 - Burst: A number of consecutive read or write operations that
655 can be queued to buffers before being flushed to memory.
656
657 - Chunk: A contiguous collection of bursts
658
659 - Transfer: A collection of chunks (be it contiguous or not)
660

3. 한국어 전문 번역

영어 원문의 문단 순서와 의미를 유지한 전체 번역입니다. 코드, 함수명, symbol과 URL은 원문 표기를 유지합니다.

Hardware 기본 원리

1-44

이 DMAengine controller documentation은 provider driver 구현을 설명합니다. 대부분의 Slave DMA controller는 같은 일반 동작 원리를 따릅니다.

controller에는 DMA transfer에 사용할 일정 수의 channel과 request line이 있습니다. request와 channel은 거의 서로 독립적입니다. channel 하나가 여러 request를 처리할 수 있습니다. 단순화하면 channel은 copy를 수행하는 entity이고 request는 어떤 endpoint가 관여하는지를 나타냅니다.

request line은 DMA를 사용할 수 있는 device에서 controller로 이어지는 물리적 line입니다. device가 transfer를 시작하려면 해당 request line을 assert해 DMA request, 즉 DRQ를 발생시킵니다.

Slave DMA request와 channel 흐름
Slave DeviceRequest LineDMA ControllerChannel
01 Wants transferAssert DRQSelect request/channelCopy between endpoints

device가 DRQ를 assert한 뒤 controller channel이 endpoint 사이 copy를 수행하는 관계를 나타냅니다.

아주 단순한 DMA controller는 transfer size 하나만 고려하며, 매 clock cycle마다 한 buffer에서 다른 buffer로 byte 하나를 옮겨 목표 크기에 도달할 때까지 반복합니다.

실제 slave device는 한 cycle에 특정 bit 수를 전송해야 할 수 있습니다. 단순 memory copy에서는 physical bus가 허용하는 만큼 옮겨 성능을 높이고 싶지만, audio device의 좁은 FIFO는 정확히 16bit 또는 24bit씩 기록해야 할 수 있습니다. 그래서 거의 모든 DMA controller는 transfer width parameter를 조절할 수 있습니다.

RAM이 source나 destination이면 일부 controller는 memory read/write를 buffer에 묶어 작은 access를 많이 수행하는 대신 큰 transfer 여러 개로 처리합니다. burst size는 controller가 더 작은 sub-transfer로 나누기 전에 연속해서 수행할 수 있는 단일 read/write 수를 정의합니다.

DMA transfer geometry
ParameterMeaningTypical Constraint
Transfer widthBits moved in one operationFIFO or bus width
Burst sizeConsecutive reads/writes before splitMemory efficiency
Transfer sizeTotal bytes requestedTransaction extent
ChunkContiguous group of burstsHardware descriptor unit

controller가 transfer를 구성할 때 쓰는 크기 단위를 구분했습니다.

Scatter-gather와 hardware descriptor collection

45-83

이론적인 단순 controller는 contiguous data block 하나만 전송할 수 있지만 실제 workload에는 non-contiguous buffer에서 contiguous buffer로 복사하는 scatter-gather가 필요합니다.

DMAEngine은 적어도 mem2dev transfer에 scatter-gather 지원을 요구합니다. controller가 직접 지원하지 않으면 driver가 software로 구현해야 하고, 고급 controller라면 hardware scatter-gather를 사용합니다.

hardware scatter-gather controller는 전송할 chunk collection으로 program합니다. transfer를 시작하면 controller가 collection을 순회해 지정된 동작을 수행합니다. collection은 보통 table 또는 linked list이며, channel에 table address와 element 수 또는 list의 첫 item을 전달합니다. DRQ가 assert되면 collection을 따라가며 data source를 찾습니다.

collection format은 hardware마다 다릅니다. 각 chunk에는 최소한 source/destination address, 각 address의 increment 여부, burst size, transfer width, transfer size가 필요합니다.

slave device는 보통 기본 상태에서 DRQ를 발생시키지 않으므로 DMA를 사용하기 전에 slave device driver에서 이를 enable해야 합니다.

지금까지는 memory-to-memory(mem2mem)와 memory-to-device(mem2dev)의 일반 원리입니다. 많은 device가 다른 transfer 또는 memory operation도 지원하며 뒤에서 설명합니다.

Linux의 DMA 지원 계층

84-99

역사적으로 DMA controller driver는 async TX API를 사용해 memory copy, XOR, cryptography 등 memory-to-memory operation을 offload했습니다.

memory-to-device transfer 수요가 생기면서 dmaengine이 확장됐습니다. 현재 async TX API는 dmaengine 위의 client layer로 구현됩니다. dmaengine은 일부 경우 이 API를 수용하며 호환성을 유지하기 위한 design choice도 포함합니다.

Async TX API의 자세한 내용은 `Documentation/crypto/async-tx-api.rst`를 참조합니다.

struct dma_device 초기화

100-144

다른 kernel framework와 마찬가지로 DMAEngine 등록은 driver가 구조체를 채워 framework에 등록하는 방식이며 여기서는 `dma_device`를 사용합니다. 일반 allocator로 구조체를 할당한 뒤 다음 field를 초기화합니다.

  • `channels`: `INIT_LIST_HEAD` 같은 macro로 list를 초기화합니다.
  • `src_addr_widths`: 지원하는 source transfer width bitmask입니다.
  • `dst_addr_widths`: 지원하는 destination transfer width bitmask입니다.
  • `directions`: mem2mem을 제외한 지원 slave direction bitmask입니다.
  • `residue_granularity`: `dma_set_residue`에 보고하는 transfer residue의 granularity입니다.
  • `dev`: 현재 driver instance와 연결된 `struct device` pointer입니다.
Residue granularity
GranularityProvider KnowledgeFramework Visibility
DescriptorNo residue detailOnly descriptor completion
SegmentTransferred chunksSegment-level progress
BurstTransferred burstsBurst-level progress

provider가 보고할 수 있는 남은 transfer 정보의 정밀도를 비교했습니다.

지원 transaction type: memory operation

145-226

다음으로 device와 driver가 지원하는 transaction type을 설정합니다. `dma_device`의 `cap_mask`를 `dma_cap_set()`으로 수정하며, capability는 `include/linux/dmaengine.h`의 `dma_transaction_type` enum에 정의되어 있습니다.

  • `DMA_MEMCPY`: memory-to-memory copy를 수행합니다. source와 destination scatter-gather list의 chunk 수와 크기가 달라도 두 list의 전체 크기 중 작은 값만큼만 전송합니다. 이는 `count`가 두 list buffer의 작은 전체 크기인 `strncpy`와 기능적으로 같습니다. modern PCI graphics card에서 host memory와 memory-mapped GPU memory 사이 pixel data를 복사하는 데 흔히 쓰이며, 큰 framebuffer를 host memory로 그대로 복사할 수 있는 OpenGL `glReadPixels()`가 대표 예입니다.
  • `DMA_XOR`: memory area에 XOR operation을 수행해 RAID5 같은 XOR 집약 작업을 가속합니다.
  • `DMA_XOR_VAL`: memory buffer에 XOR algorithm으로 parity check를 수행합니다.
  • `DMA_PQ`: RAID6 P+Q 계산을 수행합니다. P는 단순 XOR이고 Q는 Reed-Solomon algorithm입니다.
  • `DMA_PQ_VAL`: memory buffer에 RAID6 P+Q algorithm으로 parity check를 수행합니다.
  • `DMA_MEMSET`: 제공된 pattern으로 memory를 채웁니다. pattern은 single-byte signed value로 취급합니다.
  • `DMA_INTERRUPT`: periodic interrupt를 생성하는 dummy transfer를 trigger합니다. client driver는 정기적으로 DMA controller interrupt를 통해 호출될 callback을 등록할 때 사용합니다.
  • `DMA_PRIVATE`: device가 slave transfer만 지원하므로 async transfer에는 사용할 수 없습니다.
  • `DMA_ASYNC_TX`: memcpy, memset, xor, pq, xor_val, pq_val을 포함한 asynchronous memory-to-memory operation을 지원합니다. DMA engine framework가 자동으로 설정하므로 device driver가 수동으로 구성하면 안 됩니다.
Memory operation capability map
CapabilityOperationTypical Use / Rule
DMA_MEMCPYCopyGPU framebuffer and host memory
DMA_XOR / XOR_VALXOR and parity checkRAID5
DMA_PQ / PQ_VALRAID6 P+Q and parity checkXOR + Reed-Solomon
DMA_MEMSETSingle-byte pattern fillMemory initialization
DMA_INTERRUPTDummy periodic interruptRegular callback
DMA_PRIVATESlave onlyUnavailable to async TX
DMA_ASYNC_TXAsync memory operationsFramework sets automatically

기본 memory capability와 대표 용도를 요약했습니다.

지원 transaction type: slave, cyclic, interleaved와 반복

227-307
  • `DMA_SLAVE`: scatter-gather를 포함한 device-to-memory transfer를 처리합니다. mem2mem처럼 single chunk와 collection을 나누지 않고 transaction type 하나가 둘 다 처리합니다. contiguous buffer 하나는 item 하나짜리 scatterlist로 표현합니다.
  • `DMA_CYCLIC`: 마지막 item이 첫 item을 가리켜 chunk collection이 반복되는 cyclic transfer를 처리합니다. audio data를 채우는 ring buffer에 흔히 사용합니다.
  • `DMA_INTERLEAVE`: non-contiguous source에서 non-contiguous destination으로 전송합니다. non-contiguous data set에서 continuous destination으로 옮기는 `DMA_SLAVE`와 다릅니다. 보통 uncompressed 2D content 일부를 display로 직접 전송할 때 사용합니다.
  • `DMA_COMPLETION_NO_ORDER`: in-order completion을 지원하지 않습니다. operation이 out of order로 끝나면 `device_tx_status`가 `DMA_OUT_OF_ORDER`를 반환해야 합니다. 모든 cookie tracking/checking API는 무효로 취급하고 현재 dmatest polling과 호환되지 않습니다. descriptor마다 unique identifier를 제공해 completion을 추적하는 것을 권장합니다.
  • `DMA_REPEAT`: `DMA_PREP_REPEAT` flag로 끝날 때 자동 반복되며 client가 교체할 수도 있습니다. 현재 interleaved transfer에만 한정되므로 `DMA_INTERLEAVE` 없이 설정하면 안 됩니다. 다른 type 지원은 client 수요가 생길 때 추가할 수 있습니다.
  • `DMA_LOAD_EOT`: `DMA_PREP_LOAD_EOT`를 설정한 새 transfer를 queue해 end of transfer(EOT)에서 반복 transfer를 교체할 수 있습니다. end of burst 같은 다른 교체 지점은 향후 client 수요에 따라 추가할 수 있습니다.

transaction type에 따라 source/destination address 변화도 달라집니다. RAM address는 transfer마다 보통 증가하거나 감소하고 ring buffer에서는 `DMA_CYCLIC`으로 순환할 수 있습니다. FIFO 같은 device register address는 대개 고정됩니다.

Transfer capability와 address behavior
CapabilityData ShapeSpecial Rule
DMA_SLAVESG to/from contiguous endpointOne-item SG for contiguous memory
DMA_CYCLICRing of chunksLast points to first
DMA_INTERLEAVENon-contiguous to non-contiguousTypical 2D content
DMA_COMPLETION_NO_ORDERAny orderNo cookie tracking; unique IDs
DMA_REPEATRepeated interleaved transferRequires DMA_PREP_REPEAT
DMA_LOAD_EOTReplace at EOTQueue DMA_PREP_LOAD_EOT

slave, cyclic, interleaved, out-of-order와 반복 transfer의 핵심 contract를 비교했습니다.

Descriptor별 metadata 지원

308-363

일부 data movement architecture는 transaction과 연결된 metadata를 사용합니다. DMA controller는 payload와 metadata를 함께 전송합니다. metadata 자체는 DMA engine이 쓰지 않으며 peripheral에 전달하거나 peripheral에서 받은 parameter, key, vector 등을 담습니다.

DMAengine framework는 descriptor metadata를 위한 generic mechanism을 제공합니다. architecture에 따라 DMA driver가 한 mode 또는 둘 다 구현하고 client driver가 사용할 mode를 선택합니다.

  • `DESC_METADATA_CLIENT`: client driver가 buffer를 할당·제공하고 `dmaengine_desc_attach_metadata()`로 descriptor에 연결합니다. `DMA_MEM_TO_DEV`와 `DEV_MEM_TO_MEM`에서는 payload와 함께 보낼 metadata를 hardware descriptor나 tightly coupled packet으로 준비합니다. `DMA_DEV_TO_MEM`에서는 completion을 알리기 전에 metadata를 client buffer로 복사하고, 완료 뒤에는 provider가 buffer를 건드리면 안 됩니다.
  • `DESC_METADATA_ENGINE`: DMA driver가 buffer를 할당·관리합니다. client는 `dmaengine_desc_get_metadata_ptr()`로 pointer, 최대 크기, 현재 유효 크기를 얻어 직접 읽거나 쓰고 `dmaengine_desc_set_metadata_len()`으로 유효 byte 수를 알립니다. provider의 `get_metadata_ptr()`는 세 값을 반환해야 하고 `set_metadata_len()`은 client가 기록한 길이를 받습니다. `DMA_DEV_TO_MEM` client가 completion callback에서 pointer를 요청하므로 callback 전에는 descriptor를 해제하면 안 됩니다.
Provider metadata contract
ModeOwnershipMEM_TO_DEV / DEV_MEM_TO_MEMDEV_TO_MEM
DESC_METADATA_CLIENTClient bufferPrepare alongside payloadCopy before callback; stop touching after completion
DESC_METADATA_ENGINEDMA driver bufferExpose pointer and accept valid lengthKeep descriptor until callback pointer access

client-owned와 engine-owned metadata의 방향별 provider 책임을 구분했습니다.

Device operation: resource와 descriptor 준비

364-437

`dma_device`에는 실제 logic을 구현하는 function pointer도 필요하며 어떤 function을 채울지는 보고한 transaction type에 따라 달라집니다.

  • `device_alloc_chan_resources` / `device_free_chan_resources`: driver가 연결된 channel에 처음 `dma_request_channel()`을 호출하거나 마지막 `dma_release_channel()`을 호출할 때 실행됩니다. channel 사용에 필요한 resource를 할당·해제하며 sleep할 수 있습니다.
  • `device_prep_dma_*`: 등록한 capability에 대응합니다. transfer buffer나 scatterlist를 받아 hardware descriptor 또는 descriptor list를 생성합니다. interrupt context에서 호출될 수 있으므로 allocation은 sleep과 emergency pool 고갈을 피하도록 `GFP_NOWAIT`를 써야 합니다. nowait allocator 부담을 줄이기 위해 probe 때 필요한 memory를 미리 할당하는 것이 좋습니다.
  • prepare function은 transfer를 나타내는 unique `dma_async_tx_descriptor` instance를 반환해야 하며 `dma_async_tx_descriptor_init()`로 초기화할 수 있습니다.
  • descriptor의 `flags`에는 원문에 driver가 수정할 수 있는지 argument flag를 그대로 써야 하는지에 대한 TODO가 남아 있습니다. `tx_submit`은 descriptor를 pending queue에 넣어 `issue_pending` 호출을 기다리는 provider function pointer입니다.
  • submitter completion 통지는 `callback_result`로 초기화할 수 있습니다. 과거의 `callback`은 transaction status를 제공하지 않아 deprecated될 예정입니다. 전달되는 `dmaengine_result`의 `result`는 `dmaengine_tx_result`로 success 또는 error condition을 나타내고 `residue`는 지원되는 transfer의 남은 byte 수를 제공합니다.

Device operation: vector, issue와 status

438-470
  • `device_prep_peripheral_dma_vec`: `device_prep_slave_sg`와 비슷하지만 장기적으로 scatterlist를 대체할 `dma_vec` 구조체 배열 pointer를 받습니다.
  • `device_issue_pending`: pending queue의 첫 descriptor를 가져와 transfer를 시작하고 완료될 때마다 다음 transaction으로 이동합니다. interrupt context에서 호출될 수 있습니다.
  • `device_tx_status`: 전달받은 transaction descriptor의 channel 잔여 byte를 보고해야 하며 channel의 현재 active descriptor를 기준으로 삼으면 안 됩니다. `tx_state`는 NULL일 수 있고 `dma_set_residue()`로 residue를 기록합니다. cyclic transfer에서는 cyclic buffer 전체 크기만 고려합니다. in-order completion을 지원하지 않고 out of order로 끝났다면 `DMA_OUT_OF_ORDER`를 반환합니다. interrupt context에서 호출될 수 있습니다.

Device operation: configuration과 channel control

471-533
  • `device_config`: argument로 받은 `dma_slave_config`에 따라 channel을 재구성합니다. 현재 queue에 있거나 실행 중인 transfer에 동기적으로 적용하지 않고 이후 transfer에만 적용해야 합니다. 구조체의 `direction` field는 prepare function의 direction argument와 중복되어 deprecated되었습니다. slave operation에만 필수이며 memcpy에는 설정하거나 기대하면 안 됩니다. 둘 다 지원하는 driver도 slave에만 사용합니다.
  • `device_pause`: channel transfer를 즉시 pause하는 synchronous operation입니다.
  • `device_resume`: channel transfer를 즉시 resume하는 synchronous operation입니다.
  • `device_terminate_all`: channel의 pending 및 ongoing transfer를 모두 abort하며 abort된 transfer의 completion callback은 호출하지 않습니다. atomic context나 descriptor completion callback 안에서 호출될 수 있으므로 sleep하면 안 됩니다. termination은 asynchronous일 수 있어 active transfer가 완전히 멈출 때까지 기다릴 의무는 없으며 `device_synchronize`와 함께 사용합니다.
  • `device_synchronize`: channel termination을 current context와 synchronize합니다. 이전 descriptor memory에 controller가 더 이상 접근하지 않고, 이전 descriptor의 모든 completion callback이 끝났으며 새 callback이 schedule되지 않음을 보장해야 합니다. sleep할 수 있습니다.
DMA provider operation context contract
OperationContextCore Contract
alloc/free_chan_resourcesProcess contextMay sleep
device_prep_dma_*May be interruptGFP_NOWAIT; preallocate
device_issue_pending / tx_statusMay be interruptStart queue / report target descriptor
device_terminate_allAtomic or callback allowedMust not sleep; may finish asynchronously
device_synchronizeProcess contextWait for hardware and callbacks; may sleep

주요 function의 interrupt/atomic 가능 여부, sleep과 동기화 요구를 정리했습니다.

기타 note: dependency, cookie와 dma_vec

534-558

다음 내용은 별도 위치가 정해지지 않은 추가 contract입니다.

  • `dma_run_dependencies()`: async TX transfer 끝에 호출해 완료로 표시하기 전에 dependent operation이 실행되도록 합니다. slave transfer에서는 무시할 수 있습니다.
  • `dma_cookie_t`: 시간이 지날수록 증가하는 DMA transaction ID입니다. 이를 추상화하는 `virt-dma`가 도입된 뒤에는 중요성이 줄었습니다.
  • `dma_vec`: DMA address와 length를 담는 작은 구조체입니다.

DMA_CTRL_ACK와 descriptor 재사용

559-594
  • `DMA_CTRL_ACK`가 clear이면 client가 receipt를 acknowledge해 dependency chain을 설정할 기회를 얻기 전까지 provider가 descriptor를 재사용할 수 없습니다. `async_tx_ack()`로 acknowledge할 수 있습니다. flag가 set이라고 해서 descriptor를 곧바로 재사용할 수 있다는 뜻은 아닙니다.
  • `DMA_CTRL_REUSE`가 set이면 완료 뒤 descriptor를 재사용할 수 있고 provider는 이를 free하면 안 됩니다. `dmaengine_desc_set_reuse()`로 재사용을 준비하며 channel capability가 reusable descriptor를 지원할 때만 성공합니다.
  • 따라서 DMA data가 사용되지 않아 두 transfer 사이 `dma_map_sg()`와 `dma_unmap_sg()`를 생략하려는 driver는 completion 직후 transfer를 resubmit할 수 있습니다.
  • descriptor를 free하는 방법은 `dmaengine_desc_clear_reuse()`로 `DMA_CTRL_REUSE`를 clear한 뒤 마지막 transaction으로 submit하는 것, 이미 reuse flag가 set된 상태에서만 성공하는 `dmaengine_desc_free()`를 명시적으로 호출하는 것, channel을 terminate하는 것입니다.
Descriptor acknowledgement와 reuse lifecycle
State / ActionProvider May Reuse?Next Step
DMA_CTRL_ACK clearNoClient calls async_tx_ack()
DMA_CTRL_ACK setNot sufficient aloneCheck reuse contract
DMA_CTRL_REUSE setYes after completionDo not free automatically
Clear reuse + final submitFinal useFree after completion
dmaengine_desc_free()Explicit freeReuse must already be set
Terminate channelNo further useDescriptor released

ACK와 REUSE flag가 provider의 descriptor 소유권에 미치는 영향을 정리했습니다.

DMA_PREP command, repeat와 load-at-EOT flag

595-632
  • `DMA_PREP_CMD`: client가 DMA API로 전달한 data가 command data임을 controller에 알립니다. 해석은 controller별로 다르며 peripheral command, register read/write처럼 일반 data descriptor와 다른 format이 필요한 작업에 쓸 수 있습니다.
  • `DMA_PREP_REPEAT`: transfer가 끝날 때 자동 반복합니다. 같은 channel에 `DMA_PREP_LOAD_EOT`를 설정한 새 transfer가 queue되면 교체됩니다. 다음 transfer에 LOAD_EOT가 없으면 client가 모든 transfer를 terminate할 때까지 현재 transfer를 반복합니다. channel이 `DMA_REPEAT` capability를 보고할 때만 지원됩니다.
  • `DMA_PREP_LOAD_EOT`: 현재 실행 중인 transfer를 end of transfer에서 교체합니다. non-repeated transfer의 기본 behavior이므로 그런 transfer에 설정해도 차이가 없습니다. repeated transfer에서는 보통 모든 새 transfer에 설정해야 하며, 빠뜨리면 channel이 마지막 반복 transfer를 계속 실행하고 새 queue를 무시해 이전 transfer에 고정된 것처럼 보입니다. channel이 `DMA_LOAD_EOT` capability를 보고할 때만 지원됩니다.
Descriptor preparation flag
FlagEffectRequirement
DMA_PREP_CMDInterpret payload as controller-specific commandProvider-specific format
DMA_PREP_REPEATRepeat automaticallyDMA_REPEAT capability
DMA_PREP_LOAD_EOTReplace current transfer at EOTDMA_LOAD_EOT capability

command descriptor와 repeated transfer 교체 규칙을 비교했습니다.

General design note: transfer 사이 latency

633-650

많은 DMAEngine driver는 end-of-transfer interrupt를 handler에서 처리하지만, 이전 transfer가 끝난 뒤 새 transfer를 시작하는 일을 포함한 대부분의 처리를 tasklet로 미룹니다.

이 design은 transfer 사이 latency에 interrupt latency뿐 아니라 tasklet scheduling latency까지 더해 channel이 그동안 idle 상태가 되고 전체 transfer rate가 낮아집니다.

따라서 새 transfer를 tasklet에서 선택하지 말고 interrupt handler로 옮겨 피할 수 없는 idle window를 최대한 줄여야 합니다.

End-of-transfer scheduling design
DesignLatency ComponentsChannel Idle Window
Choose next in taskletInterrupt + tasklet schedulingLonger
Choose next in interrupt handlerInterrupt onlyShorter

tasklet 기반 시작과 interrupt handler 기반 시작의 idle window를 비교했습니다.

Glossary

651-659
  • Burst: memory로 flush하기 전에 buffer에 queue할 수 있는 연속 read 또는 write operation 수입니다.
  • Chunk: contiguous burst collection입니다.
  • Transfer: contiguous 여부와 관계없이 chunk collection입니다.