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