| 1 |
========================== |
| 2 |
Bulk Register Access (BRA) |
| 3 |
========================== |
| 4 |
|
| 5 |
Conventions |
| 6 |
----------- |
| 7 |
|
| 8 |
Capitalized words used in this documentation are intentional and refer |
| 9 |
to concepts of the SoundWire 1.x specification. |
| 10 |
|
| 11 |
Introduction |
| 12 |
------------ |
| 13 |
|
| 14 |
The SoundWire 1.x specification provides a mechanism to speed-up |
| 15 |
command/control transfers by reclaiming parts of the audio |
| 16 |
bandwidth. The Bulk Register Access (BRA) protocol is a standard |
| 17 |
solution based on the Bulk Payload Transport (BPT) definitions. |
| 18 |
|
| 19 |
The regular control channel uses Column 0 and can only send/retrieve |
| 20 |
one byte per frame with write/read commands. With a typical 48kHz |
| 21 |
frame rate, only 48kB/s can be transferred. |
| 22 |
|
| 23 |
The optional Bulk Register Access capability can transmit up to 12 |
| 24 |
Mbits/s and reduce transfer times by several orders of magnitude, but |
| 25 |
has multiple design constraints: |
| 26 |
|
| 27 |
(1) Each frame can only support a read or a write transfer, with a |
| 28 |
10-byte overhead per frame (header and footer response). |
| 29 |
|
| 30 |
(2) The read/writes SHALL be from/to contiguous register addresses |
| 31 |
in the same frame. A fragmented register space decreases the |
| 32 |
efficiency of the protocol by requiring multiple BRA transfers |
| 33 |
scheduled in different frames. |
| 34 |
|
| 35 |
(3) The targeted Peripheral device SHALL support the optional Data |
| 36 |
Port 0, and likewise the Manager SHALL expose audio-like Ports |
| 37 |
to insert BRA packets in the audio payload using the concepts of |
| 38 |
Sample Interval, HSTART, HSTOP, etc. |
| 39 |
|
| 40 |
(4) The BRA transport efficiency depends on the available |
| 41 |
bandwidth. If there are no on-going audio transfers, the entire |
| 42 |
frame minus Column 0 can be reclaimed for BRA. The frame shape |
| 43 |
also impacts efficiency: since Column0 cannot be used for |
| 44 |
BTP/BRA, the frame should rely on a large number of columns and |
| 45 |
minimize the number of rows. The bus clock should be as high as |
| 46 |
possible. |
| 47 |
|
| 48 |
(5) The number of bits transferred per frame SHALL be a multiple of |
| 49 |
8 bits. Padding bits SHALL be inserted if necessary at the end |
| 50 |
of the data. |
| 51 |
|
| 52 |
(6) The regular read/write commands can be issued in parallel with |
| 53 |
BRA transfers. This is convenient to e.g. deal with alerts, jack |
| 54 |
detection or change the volume during firmware download, but |
| 55 |
accessing the same address with two independent protocols has to |
| 56 |
be avoided to avoid undefined behavior. |
| 57 |
|
| 58 |
(7) Some implementations may not be capable of handling the |
| 59 |
bandwidth of the BRA protocol, e.g. in the case of a slow I2C |
| 60 |
bus behind the SoundWire IP. In this case, the transfers may |
| 61 |
need to be spaced in time or flow-controlled. |
| 62 |
|
| 63 |
(8) Each BRA packet SHALL be marked as 'Active' when valid data is |
| 64 |
to be transmitted. This allows for software to allocate a BRA |
| 65 |
stream but not transmit/discard data while processing the |
| 66 |
results or preparing the next batch of data, or allowing the |
| 67 |
peripheral to deal with the previous transfer. In addition BRA |
| 68 |
transfer can be started early on without data being ready. |
| 69 |
|
| 70 |
(9) Up to 470 bytes may be transmitted per frame. |
| 71 |
|
| 72 |
(10) The address is represented with 32 bits and does not rely on |
| 73 |
the paging registers used for the regular command/control |
| 74 |
protocol in Column 0. |
| 75 |
|
| 76 |
|
| 77 |
Error checking |
| 78 |
-------------- |
| 79 |
|
| 80 |
Firmware download is one of the key usages of the Bulk Register Access |
| 81 |
protocol. To make sure the binary data integrity is not compromised by |
| 82 |
transmission or programming errors, each BRA packet provides: |
| 83 |
|
| 84 |
(1) A CRC on the 7-byte header. This CRC helps the Peripheral Device |
| 85 |
check if it is addressed and set the start address and number of |
| 86 |
bytes. The Peripheral Device provides a response in Byte 7. |
| 87 |
|
| 88 |
(2) A CRC on the data block (header excluded). This CRC is |
| 89 |
transmitted as the last-but-one byte in the packet, prior to the |
| 90 |
footer response. |
| 91 |
|
| 92 |
The header response can be one of: |
| 93 |
(a) Ack |
| 94 |
(b) Nak |
| 95 |
(c) Not Ready |
| 96 |
|
| 97 |
The footer response can be one of: |
| 98 |
(1) Ack |
| 99 |
(2) Nak (CRC failure) |
| 100 |
(3) Good (operation completed) |
| 101 |
(4) Bad (operation failed) |
| 102 |
|
| 103 |
Example frame |
| 104 |
------------- |
| 105 |
|
| 106 |
The example below is not to scale and makes simplifying assumptions |
| 107 |
for clarity. The different chunks in the BRA packets are not required |
| 108 |
to start on a new SoundWire Row, and the scale of data may vary. |
| 109 |
|
| 110 |
:: |
| 111 |
|
| 112 |
+---+--------------------------------------------+ |
| 113 |
+ | | |
| 114 |
+ | BRA HEADER | |
| 115 |
+ | | |
| 116 |
+ +--------------------------------------------+ |
| 117 |
+ C | HEADER CRC | |
| 118 |
+ O +--------------------------------------------+ |
| 119 |
+ M | HEADER RESPONSE | |
| 120 |
+ M +--------------------------------------------+ |
| 121 |
+ A | | |
| 122 |
+ N | | |
| 123 |
+ D | DATA | |
| 124 |
+ | | |
| 125 |
+ | | |
| 126 |
+ | | |
| 127 |
+ +--------------------------------------------+ |
| 128 |
+ | DATA CRC | |
| 129 |
+ +--------------------------------------------+ |
| 130 |
+ | FOOTER RESPONSE | |
| 131 |
+---+--------------------------------------------+ |
| 132 |
|
| 133 |
|
| 134 |
Assuming the frame uses N columns, the configuration shown above can |
| 135 |
be programmed by setting the DP0 registers as: |
| 136 |
|
| 137 |
- HSTART = 1 |
| 138 |
- HSTOP = N - 1 |
| 139 |
- Sampling Interval = N |
| 140 |
- WordLength = N - 1 |
| 141 |
|
| 142 |
Addressing restrictions |
| 143 |
----------------------- |
| 144 |
|
| 145 |
The Device Number specified in the Header follows the SoundWire |
| 146 |
definitions, and broadcast and group addressing are permitted. For now |
| 147 |
the Linux implementation only allows for a single BPT transfer to a |
| 148 |
single device at a time. This might be revisited at a later point as |
| 149 |
an optimization to send the same firmware to multiple devices, but |
| 150 |
this would only be beneficial for single-link solutions. |
| 151 |
|
| 152 |
In the case of multiple Peripheral devices attached to different |
| 153 |
Managers, the broadcast and group addressing is not supported by the |
| 154 |
SoundWire specification. Each device must be handled with separate BRA |
| 155 |
streams, possibly in parallel - the links are really independent. |
| 156 |
|
| 157 |
Unsupported features |
| 158 |
-------------------- |
| 159 |
|
| 160 |
The Bulk Register Access specification provides a number of |
| 161 |
capabilities that are not supported in known implementations, such as: |
| 162 |
|
| 163 |
(1) Transfers initiated by a Peripheral Device. The BRA Initiator is |
| 164 |
always the Manager Device. |
| 165 |
|
| 166 |
(2) Flow-control capabilities and retransmission based on the |
| 167 |
'NotReady' header response require extra buffering in the |
| 168 |
SoundWire IP and are not implemented. |
| 169 |
|
| 170 |
Bi-directional handling |
| 171 |
----------------------- |
| 172 |
|
| 173 |
The BRA protocol can handle writes as well as reads, and in each |
| 174 |
packet the header and footer response are provided by the Peripheral |
| 175 |
Target device. On the Peripheral device, the BRA protocol is handled |
| 176 |
by a single DP0 data port, and at the low-level the bus ownership can |
| 177 |
will change for header/footer response as well as the data transmitted |
| 178 |
during a read. |
| 179 |
|
| 180 |
On the host side, most implementations rely on a Port-like concept, |
| 181 |
with two FIFOs consuming/generating data transfers in parallel |
| 182 |
(Host->Peripheral and Peripheral->Host). The amount of data |
| 183 |
consumed/produced by these FIFOs is not symmetrical, as a result |
| 184 |
hardware typically inserts markers to help software and hardware |
| 185 |
interpret raw data |
| 186 |
|
| 187 |
Each packet will typically have: |
| 188 |
|
| 189 |
(1) a 'Start of Packet' indicator. |
| 190 |
|
| 191 |
(2) an 'End of Packet' indicator. |
| 192 |
|
| 193 |
(3) a packet identifier to correlate the data requested and |
| 194 |
transmitted, and the error status for each frame |
| 195 |
|
| 196 |
Hardware implementations can check errors at the frame level, and |
| 197 |
retry a transfer in case of errors. However, as for the flow-control |
| 198 |
case, this requires extra buffering and intelligence in the |
| 199 |
hardware. The Linux support assumes that the entire transfer is |
| 200 |
cancelled if a single error is detected in one of the responses. |
| 201 |
|
| 202 |
Abstraction required |
| 203 |
~~~~~~~~~~~~~~~~~~~~ |
| 204 |
|
| 205 |
There are no standard registers or mandatory implementation at the |
| 206 |
Manager level, so the low-level BPT/BRA details must be hidden in |
| 207 |
Manager-specific code. For example the Cadence IP format above is not |
| 208 |
known to the codec drivers. |
| 209 |
|
| 210 |
Likewise, codec drivers should not have to know the frame size. The |
| 211 |
computation of CRC and handling of responses is handled in helpers and |
| 212 |
Manager-specific code. |
| 213 |
|
| 214 |
The host BRA driver may also have restrictions on pages allocated for |
| 215 |
DMA, or other host-DSP communication protocols. The codec driver |
| 216 |
should not be aware of any of these restrictions, since it might be |
| 217 |
reused in combination with different implementations of Manager IPs. |
| 218 |
|
| 219 |
Concurrency between BRA and regular read/write |
| 220 |
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| 221 |
|
| 222 |
The existing 'nread/nwrite' API already relies on a notion of start |
| 223 |
address and number of bytes, so it would be possible to extend this |
| 224 |
API with a 'hint' requesting BPT/BRA be used. |
| 225 |
|
| 226 |
However BRA transfers could be quite long, and the use of a single |
| 227 |
mutex for regular read/write and BRA is a show-stopper. Independent |
| 228 |
operation of the control/command and BRA transfers is a fundamental |
| 229 |
requirement, e.g. to change the volume level with the existing regmap |
| 230 |
interface while downloading firmware. The integration must however |
| 231 |
ensure that there are no concurrent access to the same address with |
| 232 |
the command/control protocol and the BRA protocol. |
| 233 |
|
| 234 |
In addition, the 'sdw_msg' structure hard-codes support for 16-bit |
| 235 |
addresses and paging registers which are irrelevant for BPT/BRA |
| 236 |
support based on native 32-bit addresses. A separate API with |
| 237 |
'sdw_bpt_msg' makes more sense. |
| 238 |
|
| 239 |
One possible strategy to speed-up all initialization tasks would be to |
| 240 |
start a BRA transfer for firmware download, then deal with all the |
| 241 |
"regular" read/writes in parallel with the command channel, and last |
| 242 |
to wait for the BRA transfers to complete. This would allow for a |
| 243 |
degree of overlap instead of a purely sequential solution. As such, |
| 244 |
the BRA API must support async transfers and expose a separate wait |
| 245 |
function. |
| 246 |
|
| 247 |
|
| 248 |
Peripheral/bus interface |
| 249 |
------------------------ |
| 250 |
|
| 251 |
The bus interface for BPT/BRA is made of two functions: |
| 252 |
|
| 253 |
- sdw_bpt_send_async(bpt_message) |
| 254 |
|
| 255 |
This function sends the data using the Manager |
| 256 |
implementation-defined capabilities (typically DMA or IPC |
| 257 |
protocol). |
| 258 |
|
| 259 |
Queueing is currently not supported, the caller |
| 260 |
needs to wait for completion of the requested transfer. |
| 261 |
|
| 262 |
- sdw_bpt_wait() |
| 263 |
|
| 264 |
This function waits for the entire message provided by the |
| 265 |
codec driver in the 'send_async' stage. Intermediate status for |
| 266 |
smaller chunks will not be provided back to the codec driver, |
| 267 |
only a return code will be provided. |
| 268 |
|
| 269 |
Regmap use |
| 270 |
~~~~~~~~~~ |
| 271 |
|
| 272 |
Existing codec drivers rely on regmap to download firmware to |
| 273 |
Peripherals. regmap exposes an async interface similar to the |
| 274 |
send/wait API suggested above, so at a high-level it would seem |
| 275 |
natural to combine BRA and regmap. The regmap layer could check if BRA |
| 276 |
is available or not, and use a regular read-write command channel in |
| 277 |
the latter case. |
| 278 |
|
| 279 |
The regmap integration will be handled in a second step. |
| 280 |
|
| 281 |
BRA stream model |
| 282 |
---------------- |
| 283 |
|
| 284 |
For regular audio transfers, the machine driver exposes a dailink |
| 285 |
connecting CPU DAI(s) and Codec DAI(s). |
| 286 |
|
| 287 |
This model is not required BRA support: |
| 288 |
|
| 289 |
(1) The SoundWire DAIs are mainly wrappers for SoundWire Data |
| 290 |
Ports, with possibly some analog or audio conversion |
| 291 |
capabilities bolted behind the Data Port. In the context of |
| 292 |
BRA, the DP0 is the destination. DP0 registers are standard and |
| 293 |
can be programmed blindly without knowing what Peripheral is |
| 294 |
connected to each link. In addition, if there are multiple |
| 295 |
Peripherals on a link and some of them do not support DP0, the |
| 296 |
write commands to program DP0 registers will generate harmless |
| 297 |
COMMAND_IGNORED responses that will be wired-ORed with |
| 298 |
responses from Peripherals which support DP0. In other words, |
| 299 |
the DP0 programming can be done with broadcast commands, and |
| 300 |
the information on the Target device can be added only in the |
| 301 |
BRA Header. |
| 302 |
|
| 303 |
(2) At the CPU level, the DAI concept is not useful for BRA; the |
| 304 |
machine driver will not create a dailink relying on DP0. The |
| 305 |
only concept that is needed is the notion of port. |
| 306 |
|
| 307 |
(3) The stream concept relies on a set of master_rt and slave_rt |
| 308 |
concepts. All of these entities represent ports and not DAIs. |
| 309 |
|
| 310 |
(4) With the assumption that a single BRA stream is used per link, |
| 311 |
that stream can connect master ports as well as all peripheral |
| 312 |
DP0 ports. |
| 313 |
|
| 314 |
(5) BRA transfers only make sense in the context of one |
| 315 |
Manager/Link, so the BRA stream handling does not rely on the |
| 316 |
concept of multi-link aggregation allowed by regular DAI links. |
| 317 |
|
| 318 |
Audio DMA support |
| 319 |
----------------- |
| 320 |
|
| 321 |
Some DMAs, such as HDaudio, require an audio format field to be |
| 322 |
set. This format is in turn used to define acceptable bursts. BPT/BRA |
| 323 |
support is not fully compatible with these definitions in that the |
| 324 |
format and bandwidth may vary between read and write commands. |
| 325 |
|
| 326 |
In addition, on Intel HDaudio Intel platforms the DMAs need to be |
| 327 |
programmed with a PCM format matching the bandwidth of the BPT/BRA |
| 328 |
transfer. The format is based on 192kHz 32-bit samples, and the number |
| 329 |
of channels varies to adjust the bandwidth. The notion of channel is |
| 330 |
completely notional since the data is not typical audio |
| 331 |
PCM. Programming such channels helps reserve enough bandwidth and adjust |
| 332 |
FIFO sizes to avoid xruns. |
| 333 |
|
| 334 |
Alignment requirements are currently not enforced at the core level |
| 335 |
but at the platform-level, e.g. for Intel the data sizes must be |
| 336 |
equal to or larger than 16 bytes. |