1.. SPDX-License-Identifier: GPL-2.0
2
3=======================
4STM32 DMA-MDMA chaining
5=======================
6
7
8Introduction
9------------
10
11  This document describes the STM32 DMA-MDMA chaining feature. But before going
12  further, let's introduce the peripherals involved.
13
14  To offload data transfers from the CPU, STM32 microprocessors (MPUs) embed
15  direct memory access controllers (DMA).
16
17  STM32MP1 SoCs embed both STM32 DMA and STM32 MDMA controllers. STM32 DMA
18  request routing capabilities are enhanced by a DMA request multiplexer
19  (STM32 DMAMUX).
20
21  **STM32 DMAMUX**
22
23  STM32 DMAMUX routes any DMA request from a given peripheral to any STM32 DMA
24  controller (STM32MP1 counts two STM32 DMA controllers) channels.
25
26  **STM32 DMA**
27
28  STM32 DMA is mainly used to implement central data buffer storage (usually in
29  the system SRAM) for different peripheral. It can access external RAMs but
30  without the ability to generate convenient burst transfer ensuring the best
31  load of the AXI.
32
33  **STM32 MDMA**
34
35  STM32 MDMA (Master DMA) is mainly used to manage direct data transfers between
36  RAM data buffers without CPU intervention. It can also be used in a
37  hierarchical structure that uses STM32 DMA as first level data buffer
38  interfaces for AHB peripherals, while the STM32 MDMA acts as a second level
39  DMA with better performance. As a AXI/AHB master, STM32 MDMA can take control
40  of the AXI/AHB bus.
41
42
43Principles
44----------
45
46  STM32 DMA-MDMA chaining feature relies on the strengths of STM32 DMA and
47  STM32 MDMA controllers.
48
49  STM32 DMA has a circular Double Buffer Mode (DBM). At each end of transaction
50  (when DMA data counter - DMA_SxNDTR - reaches 0), the memory pointers
51  (configured with DMA_SxSM0AR and DMA_SxM1AR) are swapped and the DMA data
52  counter is automatically reloaded. This allows the SW or the STM32 MDMA to
53  process one memory area while the second memory area is being filled/used by
54  the STM32 DMA transfer.
55
56  With STM32 MDMA linked-list mode, a single request initiates the data array
57  (collection of nodes) to be transferred until the linked-list pointer for the
58  channel is null. The channel transfer complete of the last node is the end of
59  transfer, unless first and last nodes are linked to each other, in such a
60  case, the linked-list loops on to create a circular MDMA transfer.
61
62  STM32 MDMA has direct connections with STM32 DMA. This enables autonomous
63  communication and synchronization between peripherals, thus saving CPU
64  resources and bus congestion. Transfer Complete signal of STM32 DMA channel
65  can triggers STM32 MDMA transfer. STM32 MDMA can clear the request generated
66  by the STM32 DMA by writing to its Interrupt Clear register (whose address is
67  stored in MDMA_CxMAR, and bit mask in MDMA_CxMDR).
68
69  .. table:: STM32 MDMA interconnect table with STM32 DMA
70
71    +--------------+----------------+-----------+------------+
72    | STM32 DMAMUX | STM32 DMA      | STM32 DMA | STM32 MDMA |
73    | channels     | channels       | Transfer  | request    |
74    |              |                | complete  |            |
75    |              |                | signal    |            |
76    +==============+================+===========+============+
77    | Channel *0*  | DMA1 channel 0 | dma1_tcf0 | *0x00*     |
78    +--------------+----------------+-----------+------------+
79    | Channel *1*  | DMA1 channel 1 | dma1_tcf1 | *0x01*     |
80    +--------------+----------------+-----------+------------+
81    | Channel *2*  | DMA1 channel 2 | dma1_tcf2 | *0x02*     |
82    +--------------+----------------+-----------+------------+
83    | Channel *3*  | DMA1 channel 3 | dma1_tcf3 | *0x03*     |
84    +--------------+----------------+-----------+------------+
85    | Channel *4*  | DMA1 channel 4 | dma1_tcf4 | *0x04*     |
86    +--------------+----------------+-----------+------------+
87    | Channel *5*  | DMA1 channel 5 | dma1_tcf5 | *0x05*     |
88    +--------------+----------------+-----------+------------+
89    | Channel *6*  | DMA1 channel 6 | dma1_tcf6 | *0x06*     |
90    +--------------+----------------+-----------+------------+
91    | Channel *7*  | DMA1 channel 7 | dma1_tcf7 | *0x07*     |
92    +--------------+----------------+-----------+------------+
93    | Channel *8*  | DMA2 channel 0 | dma2_tcf0 | *0x08*     |
94    +--------------+----------------+-----------+------------+
95    | Channel *9*  | DMA2 channel 1 | dma2_tcf1 | *0x09*     |
96    +--------------+----------------+-----------+------------+
97    | Channel *10* | DMA2 channel 2 | dma2_tcf2 | *0x0A*     |
98    +--------------+----------------+-----------+------------+
99    | Channel *11* | DMA2 channel 3 | dma2_tcf3 | *0x0B*     |
100    +--------------+----------------+-----------+------------+
101    | Channel *12* | DMA2 channel 4 | dma2_tcf4 | *0x0C*     |
102    +--------------+----------------+-----------+------------+
103    | Channel *13* | DMA2 channel 5 | dma2_tcf5 | *0x0D*     |
104    +--------------+----------------+-----------+------------+
105    | Channel *14* | DMA2 channel 6 | dma2_tcf6 | *0x0E*     |
106    +--------------+----------------+-----------+------------+
107    | Channel *15* | DMA2 channel 7 | dma2_tcf7 | *0x0F*     |
108    +--------------+----------------+-----------+------------+
109
110  STM32 DMA-MDMA chaining feature then uses a SRAM buffer. STM32MP1 SoCs embed
111  three fast access static internal RAMs of various size, used for data storage.
112  Due to STM32 DMA legacy (within microcontrollers), STM32 DMA performances are
113  bad with DDR, while they are optimal with SRAM. Hence the SRAM buffer used
114  between STM32 DMA and STM32 MDMA. This buffer is split in two equal periods
115  and STM32 DMA uses one period while STM32 MDMA uses the other period
116  simultaneously.
117  ::
118
119                    dma[1:2]-tcf[0:7]
120                   .----------------.
121     ____________ '    _________     V____________
122    | STM32 DMA  |    /  __|>_  \    | STM32 MDMA |
123    |------------|   |  /     \  |   |------------|
124    | DMA_SxM0AR |<=>| | SRAM  | |<=>| []-[]...[] |
125    | DMA_SxM1AR |   |  \_____/  |   |            |
126    |____________|    \___<|____/    |____________|
127
128  STM32 DMA-MDMA chaining uses (struct dma_slave_config).peripheral_config to
129  exchange the parameters needed to configure MDMA. These parameters are
130  gathered into a u32 array with three values:
131
132  * the STM32 MDMA request (which is actually the DMAMUX channel ID),
133  * the address of the STM32 DMA register to clear the Transfer Complete
134    interrupt flag,
135  * the mask of the Transfer Complete interrupt flag of the STM32 DMA channel.
136
137Device Tree updates for STM32 DMA-MDMA chaining support
138-------------------------------------------------------
139
140  **1. Allocate a SRAM buffer**
141
142    SRAM device tree node is defined in SoC device tree. You can refer to it in
143    your board device tree to define your SRAM pool.
144    ::
145
146          &sram {
147                  my_foo_device_dma_pool: dma-sram@0 {
148                          reg = <0x0 0x1000>;
149                  };
150          };
151
152    Be careful of the start index, in case there are other SRAM consumers.
153    Define your pool size strategically: to optimise chaining, the idea is that
154    STM32 DMA and STM32 MDMA can work simultaneously, on each buffer of the
155    SRAM.
156    If the SRAM period is greater than the expected DMA transfer, then STM32 DMA
157    and STM32 MDMA will work sequentially instead of simultaneously. It is not a
158    functional issue but it is not optimal.
159
160    Don't forget to refer to your SRAM pool in your device node. You need to
161    define a new property.
162    ::
163
164          &my_foo_device {
165                  ...
166                  my_dma_pool = &my_foo_device_dma_pool;
167          };
168
169    Then get this SRAM pool in your foo driver and allocate your SRAM buffer.
170
171  **2. Allocate a STM32 DMA channel and a STM32 MDMA channel**
172
173    You need to define an extra channel in your device tree node, in addition to
174    the one you should already have for "classic" DMA operation.
175
176    This new channel must be taken from STM32 MDMA channels, so, the phandle of
177    the DMA controller to use is the MDMA controller's one.
178    ::
179
180          &my_foo_device {
181                  [...]
182                  my_dma_pool = &my_foo_device_dma_pool;
183                  dmas = <&dmamux1 ...>,                // STM32 DMA channel
184                         <&mdma1 0 0x3 0x1200000a 0 0>; // + STM32 MDMA channel
185          };
186
187    Concerning STM32 MDMA bindings:
188
189    1. The request line number : whatever the value here, it will be overwritten
190    by MDMA driver with the STM32 DMAMUX channel ID passed through
191    (struct dma_slave_config).peripheral_config
192
193    2. The priority level : choose Very High (0x3) so that your channel will
194    take priority other the other during request arbitration
195
196    3. A 32bit mask specifying the DMA channel configuration : source and
197    destination address increment, block transfer with 128 bytes per single
198    transfer
199
200    4. The 32bit value specifying the register to be used to acknowledge the
201    request: it will be overwritten by MDMA driver, with the DMA channel
202    interrupt flag clear register address passed through
203    (struct dma_slave_config).peripheral_config
204
205    5. The 32bit mask specifying the value to be written to acknowledge the
206    request: it will be overwritten by MDMA driver, with the DMA channel
207    Transfer Complete flag passed through
208    (struct dma_slave_config).peripheral_config
209
210Driver updates for STM32 DMA-MDMA chaining support in foo driver
211----------------------------------------------------------------
212
213  **0. (optional) Refactor the original sg_table if dmaengine_prep_slave_sg()**
214
215    In case of dmaengine_prep_slave_sg(), the original sg_table can't be used as
216    is. Two new sg_tables must be created from the original one. One for
217    STM32 DMA transfer (where memory address targets now the SRAM buffer instead
218    of DDR buffer) and one for STM32 MDMA transfer (where memory address targets
219    the DDR buffer).
220
221    The new sg_list items must fit SRAM period length. Here is an example for
222    DMA_DEV_TO_MEM:
223    ::
224
225      /*
226        * Assuming sgl and nents, respectively the initial scatterlist and its
227        * length.
228        * Assuming sram_dma_buf and sram_period, respectively the memory
229        * allocated from the pool for DMA usage, and the length of the period,
230        * which is half of the sram_buf size.
231        */
232      struct sg_table new_dma_sgt, new_mdma_sgt;
233      struct scatterlist *s, *_sgl;
234      dma_addr_t ddr_dma_buf;
235      u32 new_nents = 0, len;
236      int i;
237
238      /* Count the number of entries needed */
239      for_each_sg(sgl, s, nents, i)
240              if (sg_dma_len(s) > sram_period)
241                      new_nents += DIV_ROUND_UP(sg_dma_len(s), sram_period);
242              else
243                      new_nents++;
244
245      /* Create sg table for STM32 DMA channel */
246      ret = sg_alloc_table(&new_dma_sgt, new_nents, GFP_ATOMIC);
247      if (ret)
248              dev_err(dev, "DMA sg table alloc failed\n");
249
250      for_each_sg(new_dma_sgt.sgl, s, new_dma_sgt.nents, i) {
251              _sgl = sgl;
252              sg_dma_len(s) = min(sg_dma_len(_sgl), sram_period);
253              /* Targets the beginning = first half of the sram_buf */
254              s->dma_address = sram_buf;
255              /*
256                * Targets the second half of the sram_buf
257                * for odd indexes of the item of the sg_list
258                */
259              if (i & 1)
260                      s->dma_address += sram_period;
261      }
262
263      /* Create sg table for STM32 MDMA channel */
264      ret = sg_alloc_table(&new_mdma_sgt, new_nents, GFP_ATOMIC);
265      if (ret)
266              dev_err(dev, "MDMA sg_table alloc failed\n");
267
268      _sgl = sgl;
269      len = sg_dma_len(sgl);
270      ddr_dma_buf = sg_dma_address(sgl);
271      for_each_sg(mdma_sgt.sgl, s, mdma_sgt.nents, i) {
272              size_t bytes = min_t(size_t, len, sram_period);
273
274              sg_dma_len(s) = bytes;
275              sg_dma_address(s) = ddr_dma_buf;
276              len -= bytes;
277
278              if (!len && sg_next(_sgl)) {
279                      _sgl = sg_next(_sgl);
280                      len = sg_dma_len(_sgl);
281                      ddr_dma_buf = sg_dma_address(_sgl);
282              } else {
283                      ddr_dma_buf += bytes;
284              }
285      }
286
287    Don't forget to release these new sg_tables after getting the descriptors
288    with dmaengine_prep_slave_sg().
289
290  **1. Set controller specific parameters**
291
292    First, use dmaengine_slave_config() with a struct dma_slave_config to
293    configure STM32 DMA channel. You just have to take care of DMA addresses,
294    the memory address (depending on the transfer direction) must point on your
295    SRAM buffer, and set (struct dma_slave_config).peripheral_size != 0.
296
297    STM32 DMA driver will check (struct dma_slave_config).peripheral_size to
298    determine if chaining is being used or not. If it is used, then STM32 DMA
299    driver fills (struct dma_slave_config).peripheral_config with an array of
300    three u32 : the first one containing STM32 DMAMUX channel ID, the second one
301    the channel interrupt flag clear register address, and the third one the
302    channel Transfer Complete flag mask.
303
304    Then, use dmaengine_slave_config with another struct dma_slave_config to
305    configure STM32 MDMA channel. Take care of DMA addresses, the device address
306    (depending on the transfer direction) must point on your SRAM buffer, and
307    the memory address must point to the buffer originally used for "classic"
308    DMA operation. Use the previous (struct dma_slave_config).peripheral_size
309    and .peripheral_config that have been updated by STM32 DMA driver, to set
310    (struct dma_slave_config).peripheral_size and .peripheral_config of the
311    struct dma_slave_config to configure STM32 MDMA channel.
312    ::
313
314      struct dma_slave_config dma_conf;
315      struct dma_slave_config mdma_conf;
316
317      memset(&dma_conf, 0, sizeof(dma_conf));
318      [...]
319      config.direction = DMA_DEV_TO_MEM;
320      config.dst_addr = sram_dma_buf;        // SRAM buffer
321      config.peripheral_size = 1;            // peripheral_size != 0 => chaining
322
323      dmaengine_slave_config(dma_chan, &dma_config);
324
325      memset(&mdma_conf, 0, sizeof(mdma_conf));
326      config.direction = DMA_DEV_TO_MEM;
327      mdma_conf.src_addr = sram_dma_buf;     // SRAM buffer
328      mdma_conf.dst_addr = rx_dma_buf;       // original memory buffer
329      mdma_conf.peripheral_size = dma_conf.peripheral_size;       // <- dma_conf
330      mdma_conf.peripheral_config = dma_config.peripheral_config; // <- dma_conf
331
332      dmaengine_slave_config(mdma_chan, &mdma_conf);
333
334  **2. Get a descriptor for STM32 DMA channel transaction**
335
336    In the same way you get your descriptor for your "classic" DMA operation,
337    you just have to replace the original sg_list (in case of
338    dmaengine_prep_slave_sg()) with the new sg_list using SRAM buffer, or to
339    replace the original buffer address, length and period (in case of
340    dmaengine_prep_dma_cyclic()) with the new SRAM buffer.
341
342  **3. Get a descriptor for STM32 MDMA channel transaction**
343
344    If you previously get descriptor (for STM32 DMA) with
345
346    * dmaengine_prep_slave_sg(), then use dmaengine_prep_slave_sg() for
347      STM32 MDMA;
348    * dmaengine_prep_dma_cyclic(), then use dmaengine_prep_dma_cyclic() for
349      STM32 MDMA.
350
351    Use the new sg_list using SRAM buffer (in case of dmaengine_prep_slave_sg())
352    or, depending on the transfer direction, either the original DDR buffer (in
353    case of DMA_DEV_TO_MEM) or the SRAM buffer (in case of DMA_MEM_TO_DEV), the
354    source address being previously set with dmaengine_slave_config().
355
356  **4. Submit both transactions**
357
358    Before submitting your transactions, you may need to define on which
359    descriptor you want a callback to be called at the end of the transfer
360    (dmaengine_prep_slave_sg()) or the period (dmaengine_prep_dma_cyclic()).
361    Depending on the direction, set the callback on the descriptor that finishes
362    the overal transfer:
363
364    * DMA_DEV_TO_MEM: set the callback on the "MDMA" descriptor
365    * DMA_MEM_TO_DEV: set the callback on the "DMA" descriptor
366
367    Then, submit the descriptors whatever the order, with dmaengine_tx_submit().
368
369  **5. Issue pending requests (and wait for callback notification)**
370
371  As STM32 MDMA channel transfer is triggered by STM32 DMA, you must issue
372  STM32 MDMA channel before STM32 DMA channel.
373
374  If any, your callback will be called to warn you about the end of the overal
375  transfer or the period completion.
376
377  Don't forget to terminate both channels. STM32 DMA channel is configured in
378  cyclic Double-Buffer mode so it won't be disabled by HW, you need to terminate
379  it. STM32 MDMA channel will be stopped by HW in case of sg transfer, but not
380  in case of cyclic transfer. You can terminate it whatever the kind of transfer.
381
382  **STM32 DMA-MDMA chaining DMA_MEM_TO_DEV special case**
383
384  STM32 DMA-MDMA chaining in DMA_MEM_TO_DEV is a special case. Indeed, the
385  STM32 MDMA feeds the SRAM buffer with the DDR data, and the STM32 DMA reads
386  data from SRAM buffer. So some data (the first period) have to be copied in
387  SRAM buffer when the STM32 DMA starts to read.
388
389  A trick could be pausing the STM32 DMA channel (that will raise a Transfer
390  Complete signal, triggering the STM32 MDMA channel), but the first data read
391  by the STM32 DMA could be "wrong". The proper way is to prepare the first SRAM
392  period with dmaengine_prep_dma_memcpy(). Then this first period should be
393  "removed" from the sg or the cyclic transfer.
394
395  Due to this complexity, rather use the STM32 DMA-MDMA chaining for
396  DMA_DEV_TO_MEM and keep the "classic" DMA usage for DMA_MEM_TO_DEV, unless
397  you're not afraid.
398
399Resources
400---------
401
402  Application note, datasheet and reference manual are available on ST website
403  (STM32MP1_).
404
405  Dedicated focus on three application notes (AN5224_, AN4031_ & AN5001_)
406  dealing with STM32 DMAMUX, STM32 DMA and STM32 MDMA.
407
408.. _STM32MP1: https://www.st.com/en/microcontrollers-microprocessors/stm32mp1-series.html
409.. _AN5224: https://www.st.com/resource/en/application_note/an5224-stm32-dmamux-the-dma-request-router-stmicroelectronics.pdf
410.. _AN4031: https://www.st.com/resource/en/application_note/dm00046011-using-the-stm32f2-stm32f4-and-stm32f7-series-dma-controller-stmicroelectronics.pdf
411.. _AN5001: https://www.st.com/resource/en/application_note/an5001-stm32cube-expansion-package-for-stm32h7-series-mdma-stmicroelectronics.pdf
412
413:Authors:
414
415- Amelie Delaunay <amelie.delaunay@foss.st.com>